Healthcare partner gains a deep-learning model for medical image triage

The healthcare partner relied on manual review for the initial triage of medical images, sorting them into diagnostic categories: cancer, non-malignant tumour and normal. It wanted to introduce automated classification to support that first pass. In a clinical setting, however, an automated model is only as useful as the confidence that can be placed in it. A single aggregate accuracy figure can conceal weaknesses in particular categories, and a model that performs well on the data it was trained on may not generalise. The partner therefore needed more than a working classifier: it needed a model whose behaviour was understood class by class, with clear evidence of where it performed well, where it confused categories, and whether it was overfitting to its training data.

VulcanTech built and trained a convolutional neural network that classifies medical images into three diagnostic categories: cancer, non-malignant tumour and normal. Images are preprocessed with OpenCV before classification, and the model itself is a Keras and TensorFlow CNN built from convolutional blocks with max-pooling. During training, validation accuracy was tracked alongside training accuracy to detect overfitting, with the model reaching 81.48% validation accuracy at epoch 10. Evaluation went beyond that headline figure. A full classification report sets out precision, recall and F1-score for each category, and a confusion matrix shows exactly which classes the model confuses with one another, giving the partner a transparent, category-level understanding of the model's strengths and limitations.