Quantifying imbalanced classification methods for leukemia detection.

Uncontrolled proliferation of B-lymphoblast cells is a common characterization of Acute Lymphoblastic Leukemia (ALL). B-lymphoblasts are found in large numbers in peripheral blood in malignant cases. Early detection of the cell in bone marrow is essential as the disease progresses rapidly if left untreated. However, automated classification of the cell is challenging, owing to its fine-grained variability with B-lymphoid precursor cells and imbalanced data points. Deep learning algorithms demonstrate potential for such fine-grained classification as well as suffer from the imbalanced class problem. In this paper, we explore different deep learning-based State-Of-The-Art (SOTA) approaches to tackle imbalanced classification problems. Our experiment includes input, GAN (Generative Adversarial Networks), and loss-based methods to mitigate the issue of imbalanced class on the challenging C-NMC and ALLIDB-2 dataset for leukemia detection. We have shown empirical evidence that loss-based methods outperform GAN-based and input-based methods in imbalanced classification scenarios.

Automated detection and grading of Invasive Ductal Carcinoma breast cancer using ensemble of deep learning models.

Invasive ductal carcinoma (IDC) breast cancer is a significant health concern for women all around the world and early detection of the disease may increase the survival rate in patients. Therefore, Computer-Aided Diagnosis (CAD) based systems can assist pathologists to detect the disease early. In this study, we present an ensemble model to detect IDC using DenseNet-121 and DenseNet-169 followed by test time augmentation (TTA). The model achieved a balanced accuracy of 92.70% and an F1-score of 95.70% outperforming the current state-of-the-art. Comparative analysis against various pre-trained deep learning models and preprocessing methods have been carried out. Qualitative analysis has also been conducted on the test dataset. After the detection of IDC breast cancer, it is important to grade it for further treatment. In our study, we also propose an ensemble model for the grading of IDC using the pre-trained DenseNet-121, DenseNet-201, ResNet-101v2, and ResNet-50 architectures. The model is inferred from two validation cohorts. For the patch-level classification, the model yielded an overall accuracy of 69.31%, 75.07%, 61.85%, and 60.50% on one validation cohort and 62.44%, 79.14%, 76.62%, and 71.05% on the second validation cohort for 4×, 10×, 20×, and 40× magnified images respectively. The same architecture is further validated using a different IDC dataset where it achieved an overall accuracy of 90.07%. The performance of the models on the detection and grading of IDC shows that they can be useful to help pathologists detect and grade the disease.

Automatic segmentation of blood cells from microscopic slides: A comparative analysis.

With the recent developments in deep learning, automatic cell segmentation from images of microscopic examination slides seems to be a solved problem as recent methods have achieved comparable results on existing benchmark datasets. However, most of the existing cell segmentation benchmark datasets either contain a single cell type, few instances of the cells, not publicly available. Therefore, it is unclear whether the performance improvements can generalize on more diverse datasets. In this paper, we present a large and diverse cell segmentation dataset BBBC041Seg1, which consists both of uninfected cells (i.e., red blood cells/RBCs, leukocytes) and infected cells (i.e., gametocytes, rings, trophozoites, and schizonts). Additionally, all cell types do not have equal instances, which encourages researchers to develop algorithms for learning from imbalanced classes in a few shot learning paradigm. Furthermore, we conduct a comparative study using both classical rule-based and recent deep learning state-of-the-art (SOTA) methods for automatic cell segmentation and provide them as strong baselines. We believe the introduction of BBBC041Seg will promote future research towards clinically applicable cell segmentation methods from microscopic examinations, which can be later used for downstream tasks such as detecting hematological diseases (i.e., malaria).

3C-GAN: class-consistent CycleGAN for malaria domain adaptation model.

Unpaired domain translation models with distribution matching loss such as CycleGAN are now widely being used to shift domain in medical images. However, synthesizing medical images using CycleGAN can lead to misdiagnosis of a medical condition as it might hallucinate unwanted features, especially if theres a data bias. This can potentially change the original class of the input image, which is a very serious problem. In this paper, we have introduced a modified distribution matching loss for CycleGAN to eliminate feature hallucination on the malaria dataset. In the context of the malaria dataset, unintentional feature hallucination may introduce a facet that resembles a parasite or remove the parasite after the translation. Our proposed approach has enabled us to shift the domain of the malaria dataset without the risk of changing their corresponding class. We have presented experimental evidence that our modified loss significantly reduced feature hallucination by preserving original class labels. The experimental results are better in comparison to the baseline (classic CycleGAN) that targets the translating domain. We believe that our approach will expedite the process of developing unsupervised unpaired GAN that is safe for clinical use.

A comparative analysis of deep learning architectures on high variation malaria parasite classification dataset.

Malaria, one of the leading causes of death in underdeveloped countries, is primarily diagnosed using microscopy. Computer-aided diagnosis of malaria is a challenging task owing to the fine-grained variability in the appearance of some uninfected and infected class. In this paper, we transform a malaria parasite object detection dataset into a classification dataset, making it the largest malaria classification dataset (63,645 cells), and evaluate the performance of several state-of-the-art deep neural network architectures pretrained on both natural and medical images on this new dataset. We provide detailed insights into the variation of the dataset and qualitative analysis of the results produced by the best models. We also evaluate the models using an independent test set to demonstrate the model's ability to generalize in different domains. Finally, we demonstrate the effect of conditional image synthesis on malaria parasite detection. We provide detailed insights into the influence of synthetic images for the class imbalance problem in the malaria diagnosis context.