Seeking an optimal approach for Computer-aided Diagnosis of Pulmonary Embolism.

Pulmonary Embolism (PE) represents a thrombus ("blood clot"), usually originating from a lower extremity vein, that travels to the blood vessels in the lung, causing vascular obstruction and in some patients death. This disorder is commonly diagnosed using Computed Tomography Pulmonary Angiography (CTPA). Deep learning holds great promise for the Computer-aided Diagnosis (CAD) of PE. However, numerous deep learning methods, such as Convolutional Neural Networks (CNN) and Transformer-based models, exist for a given task, causing great confusion regarding the development of CAD systems for PE. To address this confusion, we present a comprehensive analysis of competing deep learning methods applicable to PE diagnosis based on four datasets. First, we use the RSNA PE dataset, which includes (weak) slice-level and exam-level labels, for PE classification and diagnosis, respectively. At the slice level, we compare CNNs with the Vision Transformer (ViT) and the Swin Transformer. We also investigate the impact of self-supervised versus (fully) supervised ImageNet pre-training, and transfer learning over training models from scratch. Additionally, at the exam level, we compare sequence model learning with our proposed transformer-based architecture, Embedding-based ViT (E-ViT). For the second and third datasets, we utilize the CAD-PE Challenge Dataset and Ferdowsi University of Mashad's PE Dataset, where we convert (strong) clot-level masks into slice-level annotations to evaluate the optimal CNN model for slice-level PE classification. Finally, we use our in-house PE-CAD dataset, which contains (strong) clot-level masks. Here, we investigate the impact of our vessel-oriented image representations and self-supervised pre-training on PE false positive reduction at the clot level across image dimensions (2D, 2.5D, and 3D). Our experiments show that (1) transfer learning boosts performance despite differences between photographic images and CTPA scans; (2) self-supervised pre-training can surpass (fully) supervised pre-training; (3) transformer-based models demonstrate comparable performance but slower convergence compared with CNNs for slice-level PE classification; (4) model trained on the RSNA PE dataset demonstrates promising performance when tested on unseen datasets for slice-level PE classification; (5) our E-ViT framework excels in handling variable numbers of slices and outperforms sequence model learning for exam-level diagnosis; and (6) vessel-oriented image representation and self-supervised pre-training both enhance performance for PE false positive reduction across image dimensions. Our optimal approach surpasses state-of-the-art results on the RSNA PE dataset, enhancing AUC by 0.62% (slice-level) and 2.22% (exam-level). On our in-house PE-CAD dataset, 3D vessel-oriented images improve performance from 80.07% to 91.35%, a remarkable 11% gain. Codes are available at GitHub.com/JLiangLab/CAD_PE.

SegPC-2021: A challenge & dataset on segmentation of Multiple Myeloma plasma cells from microscopic images.

Multiple Myeloma (MM) is an emerging ailment of global concern. Its diagnosis at the early stages is critical for recovery. Therefore, efforts are underway to produce digital pathology tools with human-level intelligence that are efficient, scalable, accessible, and cost-effective. Following the trend, a medical imaging challenge on "Segmentation of Multiple Myeloma Plasma Cells in Microscopic Images (SegPC-2021)" was organized at the IEEE International Symposium on Biomedical Imaging (ISBI), 2021, France. The challenge addressed the problem of cell segmentation in microscopic images captured from the slides prepared from the bone marrow aspirate of patients diagnosed with Multiple Myeloma. The challenge released a total of 775 images with 690 and 85 images of sizes 2040×1536 and 1920×2560 pixels, respectively, captured from two different (microscope and camera) setups. The participants had to segment the plasma cells with a separate label on each cell's nucleus and cytoplasm. This problem comprises many challenges, including a reduced color contrast between the cytoplasm and the background, and the clustering of cells with a feeble boundary separation of individual cells. To our knowledge, the SegPC-2021 challenge dataset is the largest publicly available annotated data on plasma cell segmentation in MM so far. The challenge targets a semi-automated tool to ensure the supervision of medical experts. It was conducted for a span of five months, from November 2020 to April 2021. Initially, the data was shared with 696 people from 52 teams, of which 41 teams submitted the results of their models on the evaluation portal in the validation phase. Similarly, 20 teams qualified for the last round, of which 16 teams submitted the results in the final test phase. All the top-5 teams employed DL-based approaches, and the best mIoU obtained on the final test set of 277 microscopic images was 0.9389. All these five models have been analyzed and discussed in detail. This challenge task is a step towards the target of creating an automated MM diagnostic tool.

C-NMC: B-lineage acute lymphoblastic leukaemia: A blood cancer dataset.

Development of computer-aided cancer diagnostic tools is an active research area owing to the advancements in deep-learning domain. Such technological solutions provide affordable and easily deployable diagnostic tools. Leukaemia, or blood cancer, is one of the leading cancers causing more than 0.3 million deaths every year. In order to aid the development of such an AI-enabled tool, we collected and curated a microscopic image dataset, namely C-NMC, of more than 15000 cancer cell images at a very high resolution of B-Lineage Acute Lymphoblastic Leukaemia (B-ALL). The dataset is prepared at the subject-level and contains images of both healthy and cancer patients. So far, this is the largest (as well as curated) dataset on B-ALL cancer in the public domain. C-NMC is available at The Cancer Imaging Archive (TCIA), USA and can be helpful for the research community worldwide for the development of B-ALL cancer diagnostic tools. This dataset was utilized in an international medical imaging challenge held at ISBI 2019 conference in Venice, Italy. In this paper, we present a detailed description and challenges of this dataset. We also present benchmarking results of all the methods applied so far on this dataset.

A CNN-based unified framework utilizing projection loss in unison with label noise handling for multiple Myeloma cancer diagnosis.

Multiple Myeloma (MM) is a malignancy of plasma cells. Similar to other forms of cancer, it demands prompt diagnosis for reducing the risk of mortality. The conventional diagnostic tools are resource-intense and hence, these solutions are not easily scalable for extending their reach to the masses. Advancements in deep learning have led to rapid developments in affordable, resource optimized, easily deployable computer-assisted solutions. This work proposes a unified framework for MM diagnosis using microscopic blood cell imaging data that addresses the key challenges of inter-class visual similarity of healthy versus cancer cells and that of the label noise of the dataset. To extract class distinctive features, we propose projection loss to maximize the projection of a sample's activation on the respective class vector besides imposing orthogonality constraints on the class vectors. This projection loss is used along with the cross-entropy loss to design a dual branch architecture that helps achieve improved performance and provides scope for targeting the label noise problem. Based on this architecture, two methodologies have been proposed to correct the noisy labels. A coupling classifier has also been proposed to resolve the conflicts in the dual-branch architecture's predictions. We have utilized a large dataset of 72 subjects (26 healthy and 46 MM cancer) containing a total of 74996 images (including 34555 training cell images and 40441 test cell images). This is so far the most extensive dataset on Multiple Myeloma cancer ever reported in the literature. An ablation study has also been carried out. The proposed architecture performs best with a balanced accuracy of 94.17% on binary cell classification of healthy versus cancer in the comparative performance with ten state-of-the-art architectures. Extensive experiments on two additional publicly available datasets of two different modalities have also been utilized for analyzing the label noise handling capability of the proposed methodology. The code will be available under https://github.com/shivgahlout/CAD-MM.

Seeking an Optimal Approach for Computer-Aided Pulmonary Embolism Detection.

Pulmonary embolism (PE) represents a thrombus ("blood clot"), usually originating from a lower extremity vein, that travels to the blood vessels in the lung, causing vascular obstruction and in some patients, death. This disorder is commonly diagnosed using CT pulmonary angiography (CTPA). Deep learning holds great promise for the computer-aided CTPA diagnosis (CAD) of PE. However, numerous competing methods for a given task in the deep learning literature exist, causing great confusion regarding the development of a CAD PE system. To address this confusion, we present a comprehensive analysis of competing deep learning methods applicable to PE diagnosis using CTPA at the both image and exam levels. At the image level, we compare convolutional neural networks (CNNs) with vision transformers, and contrast self-supervised learning (SSL) with supervised learning, followed by an evaluation of transfer learning compared with training from scratch. At the exam level, we focus on comparing conventional classification (CC) with multiple instance learning (MIL). Our extensive experiments consistently show: (1) transfer learning consistently boosts performance despite differences between natural images and CT scans, (2) transfer learning with SSL surpasses its supervised counterparts; (3) CNNs outperform vision transformers, which otherwise show satisfactory performance; and (4) CC is, surprisingly, superior to MIL. Compared with the state of the art, our optimal approach provides an AUC gain of 0.2% and 1.05% for image-level and exam-level, respectively.

GCTI-SN: Geometry-inspired chemical and tissue invariant stain normalization of microscopic medical images.

Stain normalization of microscopic images is the first pre-processing step in any computer-assisted automated diagnostic tool. This paper proposes Geometry-inspired Chemical-invariant and Tissue Invariant Stain Normalization method, namely GCTI-SN, for microscopic medical images. The proposed GCTI-SN method corrects for illumination variation, stain chemical, and stain quantity variation in a unified framework by exploiting the underlying color vector space's geometry. While existing stain normalization methods have demonstrated their results on a single tissue and stain type, GCTI-SN is benchmarked on three cancer datasets of three cell/tissue types prepared with two different stain chemicals. GCTI-SN method is also benchmarked against the existing methods via quantitative and qualitative results, validating its robustness for stain chemical and cell/tissue type. Further, the utility and the efficacy of the proposed GCTI-SN stain normalization method is demonstrated diagnostically in the application of breast cancer detection via a CNN-based classifier.

SDCT-AuxNetθ: DCT augmented stain deconvolutional CNN with auxiliary classifier for cancer diagnosis.

Acute lymphoblastic leukemia (ALL) is a pervasive pediatric white blood cell cancer across the globe. With the popularity of convolutional neural networks (CNNs), computer-aided diagnosis of cancer has attracted considerable attention. Such tools are easily deployable and are cost-effective. Hence, these can enable extensive coverage of cancer diagnostic facilities. However, the development of such a tool for ALL cancer was challenging so far due to the non-availability of a large training dataset. The visual similarity between the malignant and normal cells adds to the complexity of the problem. This paper discusses the recent release of a large dataset and presents a novel deep learning architecture for the classification of cell images of ALL cancer. The proposed architecture, namely, SDCT-AuxNetθ is a 2-module framework that utilizes a compact CNN as the main classifier in one module and a Kernel SVM as the auxiliary classifier in the other one. While CNN classifier uses features through bilinear-pooling, spectral-averaged features are used by the auxiliary classifier. Further, this CNN is trained on the stain deconvolved quantity images in the optical density domain instead of the conventional RGB images. A novel test strategy is proposed that exploits both the classifiers for decision making using the confidence scores of their predicted class labels. Elaborate experiments have been carried out on our recently released public dataset of 15114 images of ALL cancer and healthy cells to establish the validity of the proposed methodology that is also robust to subject-level variability. A weighted F1 score of 94.8% is obtained that is best so far on this challenging dataset.