A study of Radiological Image-Based Bone sarcoma Detection Using Transfer Learning

Resumen

Bone sarcoma occurs primarily in children, adolescents and adults. Diagnostic assessment has traditionally involved subjective and often time-consuming assessments of imaging modalities such as X-ray, MRI and CT scans. This paper introduces a framework based on deep learning for automated classification of Bone Sarcoma from standard imaging modalities. The modernization utilizes MobileNetV2, a labeled dataset trained on ImageNet, to efficiently extract significant features while limiting computation. Preprocessing of the dataset included image normalization, resizing and augmentation using flipping, rotation, changing the zoom level and adding contrast. The dataset had a split of 80% affected and remaining 20% unaffected, respectively. During the fine-tuning phase the last layers of the model were unfrozen, and models trained at a reduced learning rate to accommodate the imaging data specific to Bone Cancer. The model trained with Adam optimizer and binary cross-entropy loss function with about 93% training accuracy and over 90% validation accuracy. Using evaluations from precision, recall, F1, and confusion matrix, the results verified the model robustness with minimal false negative rates being crucial for medical diagnostic. The results indicated that the suggested approach also provides a reliable, lightweight, and accurate diagnostic support for radiologists.