• deep learning
• oral cancer
• saliency maps
• screening
• segment-driven classification
• soft segmentation

Oral cancer is a major health problem requiring accurate screening, and medical imaging based on deep learning (DL) has proven to be an effective solution. This thesis work addresses the oral cancer classification task by employing different convolutional architectures. The goal of this work is to improve classification performance by incorporating segment information. In the proposed experiments, traditional classification training is compared with two segment-driven strategies. The first approach involves training a dedicated neural network (NN) to predict masks, which are then used to classify masked images to hide redundant information. In addition to the common hard-masking approach, an alternative relying on soft-masks to weigh the contribution of each pixel to the final classification is adopted. Then, is proposed a second approach involving the training of a NN via CrossEntropyIoU, a loss function composed of the CrossEntropy for training a classifier, and the Intersection over Union measuring the mismatch between the activation map and the mask. Experiments show implementing segment-driven strategies enhances the accuracy and the training speed. Each approach is evaluated on a dataset acquired during data collection by the medical equipment of the team. The research outcomes reveal insights into segmentation-driven DL techniques to improve oral cancer recognition.