Medical Image Analysis for Cancer Diagnosis Using Deep Learning

Abstract

Medical image analytics is essential in clinical diagnosis as they are critical for identifying various medical conditions precisely. However, routine microscopic examination of smears by pathologists/experts is time-consuming, labor-intensive, prone to error, and subjected to inter-observer variability. Computer-aided diagnostics play a crucial role in the early detection of diseases such as colorectal cancer, leukemia, etc., thereby improving long-term survival rates. In this work, deep learning-based segmentation and classification methodologies have been designed, which overcomes some of the limitations of the traditional methods. Custom-tailored U-net segmentation network was proposed for delineating the deformed gland objects from the colorectal histopathological images of malignant subjects. Multiclass classification of cell types is essential for building an automated disease detection system. Convnet feature-based classification methodology was proposed for classifying these cell types from imbalanced microscopic datasets, which were otherwise quite challenging to classify. A resilient VGG16-adapted fine-tuned feature extractor model was built to extract distinguishing features from microscopic images of single leukocyte cells. It is a critical component in building leukemia diagnostic systems. The robustness of the proposed model was found successful in interpreting the differences in a broad range of leukocyte cell types, subsequently helping achieve better accuracy and sensitivity even on a new leukocyte dataset. The last task was building an underlying classifier for detecting cancerous cells to aid cervical cancer diagnosis. The proposed work aimed to develop a robust binary classifier capable of classifying cervical cells as normal and cancerous using a modified EfficientNet model. The proposed methodology was tested as independent experiments over three public pap smear datasets and showed higher or akin performance compared to results reported in the literature. newline newline newline