American Journal for Young Scientists ยท ISSN 3068-9015
Computational Pathology for Colon Cancer: Combining Deep Feature Extraction, PCA Clustering, and Nuclear Segmentation
Authors: Noah Kaleekal
Published: October 23, 2025
Category: Biology
Abstract
Colon adenocarcinoma is among the most prevalent cancers worldwide, and an accurate histopathological diagnosis is critical for guiding treatment. However, conventional diagnoses can be limited by observer variability and the potential for false negatives. In this study, a computational framework is developed, that integrates image classification, deep feature visualization, and morphometric analysis to distinguish Hematoxylin and Eosin (H&E)-stained slides between Colon adenocarcinoma (COAD) and normal colon tissue. A ResNet18-based classifier was then trained on a balanced dataset of COAD and normal tissue samples from The Cancer Genome Atlas. The model achieved a macro-averaged F1-score of 0.9330, with perfect recall for benign samples. Deep features extracted from the network were further analyzed with Principal Component Analysis, revealing clear clustering between cancerous and normal tissues. To capture morphological differences, nuclei segmentation was applied with Cellpose. Quantitative analysis showed that COAD tissue exhibited significantly larger nuclei area fractions compared to normal tissue, consistent with the well-established phenomenon of nuclear enlargement in malignancy. Representative examples confirmed that COAD nuclei were more irregular and heterogeneous, while nuclei of benign cells remained uniform. Taken together, these results demonstrate that deep learning methods can reliably capture visual and morphometric features, distinguishing malignant from benign colon tissues. Beyond high classification accuracy, the integration of feature visualization and segmentation provides biologically interpretable insights, aligning with pathological hallmarks of cancer. This work highlights the potential for automated image analysis to complement diagnostic workflows and reduce misclassification in colon cancer pathology.