A Novel Curriculum Sampling Method for Hard Observations Classification in Breast Cancer Ultrasound

Breast cancer is the most common neoplastic disease among women, with over 2.3 million diagnoses and 685,000 deaths recorded globally in 2020. In Brazil, it is estimated that 73,610 cases will be diagnosed between 2023 and 2025, corresponding to 66.54 new cases per 100,000 women. However, the prognosis tends to be favourable if treated early. Thus, it is essential to diagnose the disease in its early stages. Ultrasound imaging helps distinguish between cystic and solid masses and determine tumour malignancy. A Computer-Aided Diagnosis (CAD) system that employs deep learning methods can serve as a second opinion for specialists, increasing the accuracy of evaluations. However, there are a few challenging cases in which the ultrasonographer makes a mistake. Thus, this work proposes a novel Curriculum Learning (CL) approach that focuses on hard cases to reduce the number of false positives and false negatives on a breast ultrasound dataset. Our proposed Adaptive Sampling approach achieves an Accuracy of 86.14 ± 3.02, a Sensitivity of 86.46 ± 2.84, a Specificity of 85.99 ± 4.00 and a F1-Score of 80.26 ± 3.67.