Training a deep neural network with a small amount of labeled data is challenging. The challenge is even more severe for medical images because of the many possible variations in the images. We propose a novel framework for few-shot chest x-ray (CXR) diagnosis. For classification problems, training with limited data may be facilitated if class-specific features can be extracted and utilized. Semantic information about the abnormalities may also be helpful in this context. To that end, we design an autoencoder-based approach that extracts visual features and decomposes them into class-agnostic and class-specific features, utilizing the semantic information of the abnormalities. The decomposition helps in efficient classification using the class-specific features. Additionally, we perform test-time adaptation to deal with possible variations in the test data compared to the training data. From this perspective, our method is one of the first of its kind. Extensive evaluations on publicly available chest x-ray datasets under few-shot settings show the effectiveness of our method. Results on the publicly available chest x-ray datasets show a 3–5% improvement in AUROC scores. Our code is available at https://github.com/mahawar2/TTAFD-FSL-CXR.git.

Federated learning methods can utilize datasets from multiple clients without requiring to share the data. Thus, these methods are helpful in preserving privacy and taking the benefit of a larger pool of data. However, most federated learning methods do not explicitly consider the importance of different layers of client models in decision-making. We propose FedImp, a federated learning method that utilizes important layers from client models for the diagnosis of breast cancer histopathology images. Our method utilizes the important layers of the client models for designing the loss function for training the client models. This may allow the client models to adapt to the data heterogeneity across clients. The important layers of client models are also utilized for aggregating the client models in the central server. Our approach limits the deviation of the client models from each other and from the aggregated model at the central server. This is likely to help in alleviating the deviation in decision-making across clients. The performance of our method is evaluated on publicly available breast cancer histopathology datasets from multiple sources. Experiments show as high as 7% increase in AUROC scores compared to state-of-the-art approaches. The code is available at https://github.com/deepMB/FedImp.

Few-shot learning (FSL) in medical image analysis presents a formidable challenge, primarily owing to the scarcity of labeled data. We propose a few-shot learning approach for the diagnosis of chest x-rays. Our method first leverages the use of an attention mechanism for better feature extraction from chest x-rays. Subsequently, we exploit auxiliary information about various abnormalities found in chest x-rays. The auxiliary information in the form of semantic signatures guides the few-shot learning process for the diagnosis of chest x-rays. We evaluate the proposed approach on multiple publicly available chest x-ray datasets. Experimental results show as high as $8 \%$ performance improvement compared to several state-of-the-art FSL approaches. Our code can be found at this link https://github.com/dpmaharathy/ICIP-ATTENTIONBASED-FEW-SHOT-DIAGNOSIS-OF-CHEST-X-RAYSUSING-SEMANTIC-SIGNATURES.git.

Breast ultrasound is useful for the diagnosis of breast tumors which can be benign or malignant. However, accurate segmentation of breast tumors and the classification of breast ultrasound into benign, malignant, or normal (no tumor) categories is challenging because of different reasons including poor contrast of the tumor region and absence of clear margins. We propose a Multibranch UNet architecture that uses multitask learning for the automated segmentation of breast tumors and classification of breast ultrasound images. Our model exploits the principle of autoencoding to achieve the aforementioned goals by utilizing salient image features. Experiments on publicly available datasets shows the superiority of our model over several state-of-the-art approaches.

Zero-shot learning (ZSL) is one of the most promising avenues of annotation-efficient machine learning. In the era of deep learning, ZSL techniques have achieved unprecedented success. However, the developments of ZSL methods have taken place mostly for natural images. ZSL for medical images has remained largely unexplored. We design a novel strategy for generalized zero-shot diagnosis of chest radiographs. In doing so, we leverage the potential of multi-view semantic embedding, a useful yet less-explored direction for ZSL. Our design also incorporates a self-training phase to tackle the problem of noisy labels alongside improving the performance for classes not seen during training. Through rigorous experiments, we show that our model trained on one dataset can produce consistent performance across test datasets from different sources including those with very different quality. Comparisons with a number of state-of-the-art techniques show the superiority of the proposed method for generalized zero-shot chest x-ray diagnosis.

We propose a shape-based solution for speckle removal from ultrasound images. The method is operable in both low contrast and high contrast imaging scenarios. The approach introduces shape information alongside structural information in the speckle removing filter. By iteratively minimizing a shape fidelity penalty to reduce speckle, the proposed filter facilitates superior segmentation of blood vessels. The effectiveness of the proposed method is established through experimentation on the ultrasound images of human blood vessels. The results show at more than 7% improvement in PSNR values compared to other state-of-the-art approaches.

We propose a fast and accurate solution to speckle reduction targeted specifically at optical coherence tomography images. The proposed speckle removing filter is designed using a novel potential function based on the gradient of the local variance of intensity. After filtering, the spatially neighboring pixels with close values of intensities converge to uniform gray values, while the edges remain intact. This filtering process results in removal of speckle without destroying the edges of the desired object. The proposed filter also prevents the generation of any false edges. Detailed experimental analysis shows at least 1-dB improvement in the peak signal-to-noise ratio for spectral domain optical coherence tomography images. The method also shows superior edge preservation, contrast, and speed compared to the state of the art in speckle removing filters.

Classification using cross-datasets (where a classifier trained using annotated image set A is used to test similar images of set B due to lack of training images in B) is important for many classification problems especially in biomedical imaging. We propose a discriminative autoencoder, useful for addressing the challenge of classification using cross-datasets. Our autoencoder learns an encoder and decoder such that the distances between the representations of the same class is minimized whereas the distances between the representations of different classes are maximized. We derive a fast algorithm to solve the aforementioned problem using the Augmented Lagrangian Alternating Directions Method of Multipliers (ADMM) approach. ADMM is a faster alternative to back-propagation which is used in standard autoencoders. The proposed method outperforms state-of-the-art representation learning tools in terms of classification results in breast cancer related histopathological image set MITOS and AMIDA and some of the benchmark image datasets.

Proportions of different phases (phase fraction) in the microstructures determine the quality of dual phase (DP) steel. So, calculation of phase fraction in the microstructures of steel samples is important for quality assurance. Manual calculation of phase fraction involves Le Pera etching of steel which is time consuming and dependent on operator efficiency. Calculation of phase fraction from Le Pera etched samples requires cumbersome manual observations. Nital etching is a faster alternative to Le Pera etching. However, due to lack of visually discriminative information, different phases cannot be identified manually from nital images. We propose a novel method for automatic calculation of phase fractions in steel microstructures from nital images using machine learning techniques. We show that regional contour patterns and local entropy (which cannot be evaluated manually) of regions of nital images are related to the formation process of the phases. We design a method that automatically evaluates regional contour patterns and local entropy from nital images of DP steel. Subsequently, we construct a random forest classifier that uses regional contour patterns and local entropy as features for classification of different phases. Our method is ∼150 times faster than manual classification. Experiments show close to 90% accuracy in classification.

We propose an improved random forest classifier that performs classification with minimum number of trees. The proposed method iteratively removes some unimportant features. Based on the number of important and unimportant features, we formulate a novel theoretical upper limit on the number of trees to be added to the forest to ensure improvement in classification accuracy. Our algorithm converges with a reduced but important set of features. We prove that further addition of trees or further reduction of features does not improve classification performance. The efficacy of the proposed approach is demonstrated through experiments on benchmark datasets. We further use the proposed classifier to detect mitotic nuclei in the histopathological datasets of breast tissues. We also apply our method on the industrial dataset of dual phase steel microstructures to classify different phases. Results of our method on different datasets show significant reduction in average classification error compared to a number of competing methods.

Grading of cancer offers insight to the occurrence and progress of the disease. The course of treatment is planned depending on the grade of cancer. Segmentation of the glandular structure of tissue is a prerequisite for grading of colon, prostate and breast cancers. Manual segmentation method is time-consuming and suffers from the curse of observer bias. We propose an automated solution for gland segmentation from hematoxylin & eosin (H&E) stained histology images. Our method relies on the biological cue rather than gland specific signatures that may vary across the slides. We construct a novel informative morphological scale space for gland segmentation. The scale space uses the entropy of the connected components in a novel manner to prevent over segmentation of objects. Our solution is fast, accurate and applicable in a clinical setup. Experiments show an average F1 score of 0.68 for 85 histology images in 20x magnification. We obtain ~ 30% improvement in F1 score compared to the area morphological scale space method.

Histopathological grading of cancer not only offers an insight to the patients' prognosis but also helps in making individual treatment plans. Mitosis counts in histopathological slides play a crucial role for invasive breast cancer grading using the Nottingham grading system. Pathologists perform this grading by manual examinations of a few thousand images for each patient. Hence, finding the mitotic figures from these images is a tedious job and also prone to observer variability due to variations in the appearances of the mitotic cells. We propose a fast and accurate approach for automatic mitosis detection from histopathological images. We employ area morphological scale space for cell segmentation. The scale space is constructed in a novel manner by restricting the scales with the maximization of relative-entropy between the cells and the background. This results in precise cell segmentation. The segmented cells are classified in mitotic and non-mitotic category using the random forest classifier. Experiments show at least 12% improvement in F1 score on more than 450 histopathological images at 40× magnification.

Muscle satellite cells, also known as myogenic precursor cells, are the dedicated stem cells responsible for postnatal skeletal muscle growth, repair, and hypertrophy. Biological studies aimed at describing satellite cell activity on their host myofiber using timelapse light microscopy enable qualitative study, but high-throughput automatic tracking of satellite cells translocating on myofibers is very difficult due to their complex motion across the three-dimensional surface of myofibers and the lack of discriminating cell features. Other complicating factors include inhomogeneous illumination, fixed focal plane, low contrast, and stage motion. We propose a semi-automated approach for satellite cell tracking on myofibers consisting of registration with illumination correction, background subtraction and particle filtering. Initial experimental results show the effectiveness of the approach.