HistAttentionNAS: Uma CNN construída via NAS para o diagnóstico do câncer do pênis usando imagens histopatológicas

Abstract

Penile cancer (PC) is a rare neoplasm which, although infrequent in developed countries, has an alarming incidence in some areas of developing countries. Early diagnosis is essential to improve prognosis and reduce the need for mutilating surgical interventions. Histopathological analysis remains the gold standard for diagnosis, but it involves subjective processes that can delay the report and lead to disagreements between specialists. In this sense, deep learning has shown promise in speeding up and standardizing assessments, helping pathologists to identify relevant patterns. This paper proposes a methodology for classifying PC in digital histopathology images. It begins by preparing the data, balancing the classes using geometric transformations and applying different pre-processing and colour disturbance techniques. Next, a Neural Architecture Search process is used to automatically define the best combination of stem blocks and attention mechanisms in the backbone stages of the network, optimizing the extraction of textures and discriminative information. In the experiments, the optimized architectures showed F1-Score values of 0.927 at 40x and 0.958 at 100x, confirming the strategy’s effectiveness. Comparisons with related work on the same basis show competitive results, particularly at 100x magnification, even without using large external bases or transfer learning. It is concluded that class balancing, pre-processing configuration and increased representativeness, combined with automation via neural architecture search (NAS), boost the performance of penile cancer classification.