Multi-task learning for segmentation and classification of tumors in 3D automated breast ultrasound images

Tumor classification and segmentation are two important tasks for computer-aided diagnosis (CAD) using 3D automated breast ultrasound (ABUS) images. However, they are challenging due to the significant shape variation of breast tumors and the fuzzy nature of ultrasound images (e.g., low contrast and signal to noise ratio). Considering the correlation between tumor classification and segmentation, we argue that learning these two tasks jointly is able to improve the outcomes of both tasks. In this paper, we propose a novel multi-task learning framework for joint segmentation and classification of tumors in ABUS images. The proposed framework consists of two sub-networks: an encoder-decoder network for segmentation and a light-weight multi-scale network for classification. To account for the fuzzy boundaries of tumors in ABUS images, our framework uses an iterative training strategy to refine feature maps with the help of probability maps obtained from previous iterations. Experimental results based on a clinical dataset of 170 3D ABUS volumes collected from 107 patients indicate that the proposed multi-task framework improves tumor segmentation and classification over the single-task learning counterparts.     

Towards bi-directional skip connections in encoder-decoder architectures and beyond

U-Net, as an encoder-decoder architecture with forward skip connections, has achieved promising results in various medical image analysis tasks. Many recent approaches have also extended U-Net with more complex building blocks, which typically increase the number of network parameters considerably. Such complexity makes the inference stage highly inefficient for clinical applications. Towards an effective yet economic segmentation network design, in this work, we propose backward skip connections that bring decoded features back to the encoder. Our design can be jointly adopted with forward skip connections in any encoder-decoder architecture forming a recurrence structure without introducing extra parameters. With the backward skip connections, we propose a U-Net based network family, namely Bi-directional O-shape networks, which set new benchmarks on multiple public medical imaging segmentation datasets. On the other hand, with the most plain architecture (BiO-Net), network computations inevitably increase along with the pre-set recurrence time. We have thus studied the deficiency bottleneck of such recurrent design and propose a novel two-phase Neural Architecture Search (NAS) algorithm, namely BiX-NAS, to search for the best multi-scale bi-directional skip connections. The ineffective skip connections are then discarded to reduce computational costs and speed up network inference. The finally searched BiX-Net yields the least network complexity and outperforms other state-of-the-art counterparts by large margins. We evaluate our methods on both 2D and 3D segmentation tasks in a total of six datasets. Extensive ablation studies have also been conducted to provide a comprehensive analysis for our proposed methods.     

Gumbel-Softmax based Neural Architecture Search for Hierarchical Brain Networks Decomposition

Understanding the brain's functional architecture has been an important topic in the neuroimaging field. A variety of brain network modeling methods have been proposed. Recently, deep neural network-based methods have shown a great advantage in modeling the hierarchical and complex functional brain networks (FBNs). However, most of these deep neural networks were handcrafted, making it time-consuming to find the relatively optimal architecture. To address this problem, we propose a novel unsupervised differentiable neural architecture search (NAS) algorithm, named Gumbel-Softmax based Neural Architecture Search (GS-NAS), to automate the architecture design of deep belief network (DBN) for hierarchical FBN decomposition. Specifically, we introduce the Gumbel-Softmax scheme to reframe the discrete architecture sampling procedure during NAS to be continuous. Guided by the reconstruction error minimization procedure, the architecture search can be driven by the intrinsic functional architecture of the brain, thereby revealing the possible hierarchical functional brain organization via DBN structure. The proposed GS-NAS algorithm can simultaneously optimize the number of hidden units for each layer and the network depth. Extensive experiment results on both task and resting-state functional magnetic resonance imaging data have demonstrated the effectiveness and efficiency of the proposed GS-NAS model. The identified hierarchically organized FBNs provide novel insight into understanding human brain function.     

Diagnostic Performance of Automated Breast Ultrasound as a Replacement for a Hand-Held Second-Look Ultrasound for Breast Lesions Detected Initially on Magnetic Resonance Imaging

To evaluate the diagnostic performance of automated breast ultrasound (ABUS) after breast magnetic resonance imaging (MRI) as a replacement for hand-held second-look ultrasound (HH-SLUS), we evaluated 58 consecutive patients with breast cancer who had additional suspicious lesions on breast MRI. All patients underwent HH-SLUS and ABUS. Three breast radiologists evaluated the detectability, location, characteristics and conspicuity of lesions on ABUS. We also evaluated inter-observer variability and compared the results with HH-SLUS results. Eighty additional suspicious lesions were identified on breast MRI. Fifteen of the 80 lesions (19%) were not detected on HH-SLUS. Eight of the 15 lesions (53%) were detected on ABUS, whereas the remaining 7 were not detected on ABUS. Among the 65 lesions detected on HH-SLUS, only 3 lesions were not detected on ABUS. The intra-class correlation coefficients for lesion location and size all exceeded 0.70, indicating high reliability. Moderate to fair agreement was found for mass shape, orientation, margin and Breast Imaging Reporting and Data System (BI-RADS) final assessment. Therefore, ABUS can reliably detect additional suspicious lesions identified on breast MRI and may help in the decision on biopsy guidance method (US vs. MRI) as a replacement tool for HH-SLUS.     

Global-Local attention network with multi-task uncertainty loss for abnormal lymph node detection in MR images

Accurate and reliable detection of abnormal lymph nodes in magnetic resonance (MR) images is very helpful for the diagnosis and treatment of numerous diseases. However, it is still a challenging task due to similar appearances between abnormal lymph nodes and other tissues. In this paper, we propose a novel network based on an improved Mask R-CNN framework for the detection of abnormal lymph nodes in MR images. Instead of laboriously collecting large-scale pixel-wise annotated training data, pseudo masks generated from RECIST bookmarks on hand are utilized as the supervision. Different from the standard Mask R-CNN architecture, there are two main innovations in our proposed network: 1) global-local attention which encodes the global and local scale context for detection and utilizes the channel attention mechanism to extract more discriminative features and 2) multi-task uncertainty loss which adaptively weights multiple objective loss functions based on the uncertainty of each task to automatically search the optimal solution. For the experiments, we built a new abnormal lymph node dataset with 821 RECIST bookmarks of 41 different types of abnormal abdominal lymph nodes from 584 different patients. The experimental results showed the superior performance of our algorithm over compared state-of-the-art approaches.     

Breast Cancer Classification in Automated Breast Ultrasound Using Multiview Convolutional Neural Network with Transfer Learning

To assist radiologists in breast cancer classification in automated breast ultrasound (ABUS) imaging, we propose a computer-aided diagnosis based on a convolutional neural network (CNN) that classifies breast lesions as benign and malignant. The proposed CNN adopts a modified Inception-v3 architecture to provide efficient feature extraction in ABUS imaging. Because the ABUS images can be visualized in transverse and coronal views, the proposed CNN provides an efficient way to extract multiview features from both views. The proposed CNN was trained and evaluated on 316 breast lesions (135 malignant and 181 benign). An observer performance test was conducted to compare five human reviewers' diagnostic performance before and after referring to the predicting outcomes of the proposed CNN. Our method achieved an area under the curve (AUC) value of 0.9468 with five-folder cross-validation, for which the sensitivity and specificity were 0.886 and 0.876, respectively. Compared with conventional machine learning-based feature extraction schemes, particularly principal component analysis (PCA) and histogram of oriented gradients (HOG), our method achieved a significant improvement in classification performance. The proposed CNN achieved a >10% increased AUC value compared with PCA and HOG. During the observer performance test, the diagnostic results of all human reviewers had increased AUC values and sensitivities after referring to the classification results of the proposed CNN, and four of the five human reviewers’ AUCs were significantly improved. The proposed CNN employing a multiview strategy showed promise for the diagnosis of breast cancer, and could be used as a second reviewer for increasing diagnostic reliability.     

Global guidance network for breast lesion segmentation in ultrasound images

Automatic breast lesion segmentation in ultrasound helps to diagnose breast cancer, which is one of the dreadful diseases that affect women globally. Segmenting breast regions accurately from ultrasound image is a challenging task due to the inherent speckle artifacts, blurry breast lesion boundaries, and inhomogeneous intensity distributions inside the breast lesion regions. Recently, convolutional neural networks (CNNs) have demonstrated remarkable results in medical image segmentation tasks. However, the convolutional operations in a CNN often focus on local regions, which suffer from limited capabilities in capturing long-range dependencies of the input ultrasound image, resulting in degraded breast lesion segmentation accuracy. In this paper, we develop a deep convolutional neural network equipped with a global guidance block (GGB) and breast lesion boundary detection (BD) modules for boosting the breast ultrasound lesion segmentation. The GGB utilizes the multi-layer integrated feature map as a guidance information to learn the long-range non-local dependencies from both spatial and channel domains. The BD modules learn additional breast lesion boundary map to enhance the boundary quality of a segmentation result refinement. Experimental results on a public dataset and a collected dataset show that our network outperforms other medical image segmentation methods and the recent semantic segmentation methods on breast ultrasound lesion segmentation. Moreover, we also show the application of our network on the ultrasound prostate segmentation, in which our method better identifies prostate regions than state-of-the-art networks.     

Learning to segment images with classification labels

Two of the most common tasks in medical imaging are classification and segmentation. Either task requires labeled data annotated by experts, which is scarce and expensive to collect. Annotating data for segmentation is generally considered to be more laborious as the annotator has to draw around the boundaries of regions of interest, as opposed to assigning image patches a class label. Furthermore, in tasks such as breast cancer histopathology, any realistic clinical application often includes working with whole slide images, whereas most publicly available training data are in the form of image patches, which are given a class label. We propose an architecture that can alleviate the requirements for segmentation-level ground truth by making use of image-level labels to reduce the amount of time spent on data curation. In addition, this architecture can help unlock the potential of previously acquired image-level datasets on segmentation tasks by annotating a small number of regions of interest. In our experiments, we show using only one segmentation-level annotation per class, we can achieve performance comparable to a fully annotated dataset.     

Deformable mapping using biomechanical models to relate corresponding lesions in digital breast tomosynthesis and automated breast ultrasound images

This work investigates the application of a deformable localization/mapping method to register lesions between the digital breast tomosynthesis (DBT) craniocaudal (CC) and mediolateral oblique (MLO) views and automated breast ultrasound (ABUS) images. This method was initially validated using compressible breast phantoms. This methodology was applied to 7 patient data sets containing 9 lesions. The automated deformable mapping algorithm uses finite element modeling and analysis to determine corresponding lesions based on the distance between their centers of mass (d_COM) in the deformed DBT model and the reference ABUS model. This technique shows that location information based on external fiducial markers is helpful in the improvement of registration results. However, use of external markers are not required for deformable registration results described by this methodology. For DBT (CC view) mapped to ABUS, the mean d_COM was 14.9 ± 6.8 mm based on 9 lesions using 6 markers in deformable analysis. For DBT (MLO view) mapped to ABUS, the mean d_COM was 13.7 ± 6.8 mm based on 8 lesions using 6 markers in analysis. Both DBT views registered to ABUS lesions showed statistically significant improvements (p ≤ 0.05) in registration using the deformable technique in comparison to a rigid registration. Application of this methodology could help improve a radiologist's characterization and accuracy in relating corresponding lesions between DBT and ABUS image datasets, especially for cases of high breast densities and multiple masses.     

Computer-aided diagnosis for the classification of breast masses in automated whole breast ultrasound images

New automated whole breast ultrasound (ABUS) machines have recently been developed and the ultrasound (US) volume dataset of the whole breast can be acquired in a standard manner. The purpose of this study was to develop a novel computer-aided diagnosis system for classification of breast masses in ABUS images. One hundred forty-seven cases (76 benign and 71 malignant breast masses) were obtained by a commercially available ABUS system. Because the distance of neighboring slices in ABUS images is fixed and small, these continuous slices were used for reconstruction as three-dimensional (3-D) US images. The 3-D tumor contour was segmented using the level-set segmentation method. Then, the 3-D features, including the texture, shape and ellipsoid fitting were extracted based on the segmented 3-D tumor contour to classify benign and malignant tumors based on the logistic regression model. The Student’s t test, Mann-Whitney U test and receiver operating characteristic (ROC) curve analysis were used for statistical analysis. From the Az values of ROC curves, the shape features (0.9138) are better than the texture features (0.8603) and the ellipsoid fitting features (0.8496) for classification. The difference was significant between shape and ellipsoid fitting features (p = 0.0382). However, combination of ellipsoid fitting features and shape features can achieve a best performance with accuracy of 85.0% (125/147), sensitivity of 84.5% (60/71), specificity of 85.5% (65/76) and the area under the ROC curve Az of 0.9466. The results showed that ABUS images could be used for computer-aided feature extraction and classification of breast tumors. (E-mail: rfchang@csie.ntu.edu.tw)     

Road extraction from aerial image data via multiple features integrated with convolution long short time memory unit network

ABSTRACT

Semantic segmentation models based on deep learning have shown remarkable performance in road extraction from high-resolution aerial images. However, it is still a difficult task to segment multiscale roads with high completeness and accuracy from complex backgrounds. To deal with this problem, this letter proposes an end to end network named Multiple features integrated with convolutional long-short time memory unit network (MFI-CLSTMN). First, in MFI-CLSTMN, the ConvLSTM unit is designed to explore and integrate the sequential correlations among features, which can alleviate the feature loss caused by the max-pooling operation. Second, the structure of dense concatenation and multiscale up-sampling combines detailed features with semantic information to preserve road details. At last, at the optimization stage, a self-adaptive composite loss function is added to handle class imbalance, such that MFI-CLSTMN can effectively train hard examples and avoids local optimum. Experiments demonstrate that MFI-CLSTMN has higher segmentation accuracy and lower computational complexity than four comparative state-of-the-art models in a consistent environment. Moreover, MFI-CLSTMN can especially protect road segmentation from netsplit and brokenness, which is hard for other models to achieve.     

Joint optimization of segmentation and shape prior from level-set-based statistical shape model,and its application to the automated segmentation of abdominal organs

The goal of this study is to provide a theoretical framework for accurately optimizing the segmentation energy considering all of the possible shapes generated from the level-set-based statistical shape model (SSM). The proposed algorithm solves the well-known open problem, in which a shape prior may not be optimal in terms of an objective functional that needs to be minimized during segmentation. The algorithm allows the selection of an optimal shape prior from among all possible shapes generated from an SSM by conducting a branch-and-bound search over an eigenshape space. The proposed algorithm does not require predefined shape templates or the construction of a hierarchical clustering tree before graph-cut segmentation. It jointly optimizes an objective functional in terms of both the shape prior and segmentation labeling, and finds an optimal solution by considering all possible shapes generated from an SSM. We apply the proposed algorithm to both pancreas and spleen segmentation using multiphase computed tomography volumes, and we compare the results obtained with those produced by a conventional algorithm employing a branch-and-bound search over a search tree of predefined shapes, which were sampled discretely from an SSM. The proposed algorithm significantly improves the segmentation performance in terms of the Jaccard index and Dice similarity index. In addition, we compare the results with the state-of-the-art multiple abdominal organs segmentation algorithm, and confirmed that the performances of both algorithms are comparable to each other. We discuss the high computational efficiency of the proposed algorithm, which was determined experimentally using a normalized number of traversed nodes in a search tree, and the extensibility of the proposed algorithm to other SSMs or energy functionals.     

Deformable Mapping Method to Relate Lesions in Dedicated Breast CT Images to Those in Automated Breast Ultrasound and Digital Breast Tomosynthesis Images

This work demonstrates the potential for using a deformable mapping method to register lesions between dedicated breast computed tomography (bCT) and both automated breast ultrasound (ABUS) and digital breast tomosynthesis (DBT) images (craniocaudal [CC] and mediolateral oblique [MLO] views). Two multi-modality breast phantoms with external fiducial markers attached were imaged by the three modalities. The DBT MLO view was excluded for the second phantom. The automated deformable mapping algorithm uses biomechanical modeling to determine corresponding lesions based on distances between their centers of mass (d_COM) in the deformed bCT model and the reference model (DBT or ABUS). For bCT to ABUS, the mean d_COM was 5.2 ± 2.6 mm. For bCT to DBT (CC), the mean d_COM was 5.1 ± 2.4 mm. For bCT to DBT (MLO), the mean d_COM was 4.7 ± 2.5 mm. This application could help improve a radiologist's efficiency and accuracy in breast lesion characterization, using multiple imaging modalities.     

Automatic segmentation of airway tree based on local intensity filter and machine learning technique in 3D chest CT volume

Purpose

Airway segmentation plays an important role in analyzing chest computed tomography (CT) volumes for computerized lung cancer detection, emphysema diagnosis and pre- and intra-operative bronchoscope navigation. However, obtaining a complete 3D airway tree structure from a CT volume is quite a challenging task. Several researchers have proposed automated airway segmentation algorithms basically based on region growing and machine learning techniques. However, these methods fail to detect the peripheral bronchial branches, which results in a large amount of leakage. This paper presents a novel approach for more accurate extraction of the complex airway tree.

Methods

This proposed segmentation method is composed of three steps. First, Hessian analysis is utilized to enhance the tube-like structure in CT volumes; then, an adaptive multiscale cavity enhancement filter is employed to detect the cavity-like structure with different radii. In the second step, support vector machine learning will be utilized to remove the false positive (FP) regions from the result obtained in the previous step. Finally, the graph-cut algorithm is used to refine the candidate voxels to form an integrated airway tree.

Results

A test dataset including 50 standard-dose chest CT volumes was used for evaluating our proposed method. The average extraction rate was about 79.1 % with the significantly decreased FP rate.

Conclusion

A new method of airway segmentation based on local intensity structure and machine learning technique was developed. The method was shown to be feasible for airway segmentation in a computer-aided diagnosis system for a lung and bronchoscope guidance system.

     

Thyroid nodule segmentation and classification in ultrasound images through intra- and inter-task consistent learning

Thyroid nodule segmentation and classification in ultrasound images are two essential but challenging tasks for computer-aided diagnosis of thyroid nodules. Since these two tasks are inherently related to each other and sharing some common features, solving them jointly with multi-task leaning is a promising direction. However, both previous studies and our experimental results confirm the problem of inconsistent predictions among these related tasks. In this paper, we summarize two types of task inconsistency according to the relationship among different tasks: intra-task inconsistency between homogeneous tasks (e.g., both tasks are pixel-wise segmentation tasks); and inter-task inconsistency between heterogeneous tasks (e.g., pixel-wise segmentation task and categorical classification task). To address the task inconsistency problems, we propose intra- and inter-task consistent learning on top of the designed multi-stage and multi-task learning network to enforce the network learn consistent predictions for all the tasks during network training. Our experimental results based on a large clinical thyroid ultrasound image dataset indicate that the proposed intra- and inter-task consistent learning can effectively eliminate both types of task inconsistency and thus improve the performance of all tasks for thyroid nodule segmentation and classification.     

Unified generative adversarial networks for multimodal segmentation from unpaired 3D medical images

To fully define the target objects of interest in clinical diagnosis, many deep convolution neural networks (CNNs) use multimodal paired registered images as inputs for segmentation tasks. However, these paired images are difficult to obtain in some cases. Furthermore, the CNNs trained on one specific modality may fail on others for images acquired with different imaging protocols and scanners. Therefore, developing a unified model that can segment the target objects from unpaired multiple modalities is significant for many clinical applications. In this work, we propose a 3D unified generative adversarial network, which unifies the any-to-any modality translation and multimodal segmentation in a single network. Since the anatomical structure is preserved during modality translation, the auxiliary translation task is used to extract the modality-invariant features and generate the additional training data implicitly. To fully utilize the segmentation-related features, we add a cross-task skip connection with feature recalibration from the translation decoder to the segmentation decoder. Experiments on abdominal organ segmentation and brain tumor segmentation indicate that our method outperforms the existing unified methods.     

Automated left ventricular segmentation from cardiac magnetic resonance images via adversarial learning with multi-stage pose estimation network and co-discriminator

Left ventricular (LV) segmentation is essential for the early diagnosis of cardiovascular diseases, which has been reported as the leading cause of death all over the world. However, automated LV segmentation from cardiac magnetic resonance images (CMRI) using the traditional convolutional neural networks (CNNs) is still a challenging task due to the limited labeled CMRI data and low tolerances to irregular scales, shapes and deformations of LV. In this paper, we propose an automated LV segmentation method based on adversarial learning by integrating a multi-stage pose estimation network (MSPN) and a co-discrimination network. Different from existing CNNs, we use a MSPN with multi-scale dilated convolution (MDC) modules to enhance the ranges of receptive field for deep feature extraction. To fully utilize both labeled and unlabeled CMRI data, we propose a novel generative adversarial network (GAN) framework for LV segmentation by combining MSPN with co-discrimination networks. Specifically, the labeled CMRI are first used to initialize our segmentation network (MSPN) and co-discrimination network. Our GAN training includes two different kinds of epochs fed with both labeled and unlabeled CMRI data alternatively, which are different from the traditional CNNs only relied on the limited labeled samples to train the segmentation networks. As both ground truth and unlabeled samples are involved in guiding training, our method not only can converge faster but also obtain a better performance in LV segmentation. Our method is evaluated using MICCAI 2009 and 2017 challenge databases. Experimental results show that our method has obtained promising performance in LV segmentation, which also outperforms the state-of-the-art methods in terms of LV segmentation accuracy from the comparison results.     

Fully convolutional multi-scale residual DenseNets for cardiac segmentation and automated cardiac diagnosis using ensemble of classifiers

Deep fully convolutional neural network (FCN) based architectures have shown great potential in medical image segmentation. However, such architectures usually have millions of parameters and inadequate number of training samples leading to over-fitting and poor generalization. In this paper, we present a novel DenseNet based FCN architecture for cardiac segmentation which is parameter and memory efficient. We propose a novel up-sampling path which incorporates long skip and short-cut connections to overcome the feature map explosion in conventional FCN based architectures. In order to process the input images at multiple scales and view points simultaneously, we propose to incorporate Inception module’s parallel structures. We propose a novel dual loss function whose weighting scheme allows to combine advantages of cross-entropy and Dice loss leading to qualitative improvements in segmentation. We demonstrate computational efficacy of incorporating conventional computer vision techniques for region of interest detection in an end-to-end deep learning based segmentation framework. From the segmentation maps we extract clinically relevant cardiac parameters and hand-craft features which reflect the clinical diagnostic analysis and train an ensemble system for cardiac disease classification. We validate our proposed network architecture on three publicly available datasets, namely: (i) Automated Cardiac Diagnosis Challenge (ACDC-2017), (ii) Left Ventricular segmentation challenge (LV-2011), (iii) 2015 Kaggle Data Science Bowl cardiac challenge data. Our approach in ACDC-2017 challenge stood second place for segmentation and first place in automated cardiac disease diagnosis tasks with an accuracy of 100% on a limited testing set (n=50). In the LV-2011 challenge our approach attained 0.74 Jaccard index, which is so far the highest published result in fully automated algorithms. In the Kaggle challenge our approach for LV volume gave a Continuous Ranked Probability Score (CRPS) of 0.0127, which would have placed us tenth in the original challenge. Our approach combined both cardiac segmentation and disease diagnosis into a fully automated framework which is computationally efficient and hence has the potential to be incorporated in computer-aided diagnosis (CAD) tools for clinical application.     

Multiscale 3D shape representation and segmentation with applications to hippocampal/caudate extraction from brain MRI

Extracting structure of interest from medical images is an important yet tedious work. Due to the image quality, the shape knowledge is widely used for assisting and constraining the segmentation process. In many previous works, shape knowledge was incorporated by first constructing a shape space from training cases, and then constraining the segmentation process to be within the learned shape space. However, such an approach has certain limitations due to the number of variations, eigen-shapemodes, that can be captured in the learned shape space. Moreover, small scale shape variances are usually overwhelmed by those in the large scale, and therefore the local shape information is lost. In this work, we present a multiscale representation for shapes with arbitrary topology, and a fully automatic method to segment the target organ/tissue from medical images using such multiscale shape information and local image features. First, we handle the problem of lacking eigen-shapemodes by providing a multiscale shape representation using the wavelet transform. Consequently, the shape variances existing in the training shapes captured by the statistical learning step are also represented at various scales. Note that by doing so, one can greatly enrich the eigen-shapemodes as well as capture small scale shape changes. Furthermore, in order to make full use of the training information, not only the shape but also the grayscale training images are utilized in a multi-atlas initialization procedure. By combining such initialization with the multiscale shape knowledge, we perform segmentation tests for challenging medical data sets where the target objects have low contrast and sharp corner structures, and demonstrate the statistically significant improvement obtained by employing such multiscale representation, in representing shapes as well as the overall shape based segmentation tasks.     

NAS-optimized topology-preserving transfer learning for differentiating cortical folding patterns

Increasing evidence suggests that cortical folding patterns of human cerebral cortex manifest overt structural and functional differences. However, for interpretability, few studies leverage advanced techniques (e.g., deep learning) to investigate the difference among cortical folds, resulting in more differences yet to be extensively explored. To this end, we proposed an effective topology-preserving transfer learning framework to differentiate cortical fMRI time series extracted from cortical folds. Our framework consists of three main parts: (1) Neural architecture search (NAS), which is used to devise a well-performing network structure based on an initialized hand-designed super-graph in an image dataset; (2) Topology-preserving transfer, which takes the model searched by NAS as the source network, keeping the topological connectivity in the network unchanged, while transforming all 2D operations including convolution and pooling into 1D, therefore resulting in a topology-preserving network, named TPNAS-Net; (3) Classification and correlation analysis, which involves using the TPNAS-Net to classify 1D cortical fMRI time series for each individual brain, and performing a group difference analysis between autism spectrum disorder (ASD) and healthy control (HC) and correlation analysis with clinical information (i.e., age). Extensive experiments on two ASD datasets obtain consistent results, demonstrating that the TPNAS-Net not only discriminates cortical folding patterns at high classification accuracy, but also captures subtle differences between ASD and HC (p-value = 0.042). In addition, we discover that there is a positive correlation between the classification accuracy and age in ASD (r = 0.39, p-value = 0.04). These findings together suggest that structural and functional differences in cortical folding patterns between ASD and HC may provide a potentially useful biomarker for the diagnosis of ASD.