Breast cancer has surpassed lung cancer as the most diagnosed in the world, and it is the fifth cancer in number of deaths. In Brazil, it is estimated that 73,610 cases will be diagnosed from 2023 to 2025, which means 66.54 new cases per 100,000 women. Nevertheless, the prognosis is good for most cases if treated early and correctly. Ultrasound images are suitable for distinguishing between cystic and solid masses and determining the malignancy of tumours through shapes, margins, echo patterns, and posterior characteristics. Thus, Computer-Aided Diagnosis systems can aid the specialist in finding and extracting nodules in ultrasound images to decide if a biopsy is necessary. However, there are rase cases with complex abnormal lesions that even the specialist may be in doubt to define a diagnosis. This rarity can cause an imbalance in training and in the efficiency of a Computer-Aided Diagnosis system. These cases would be the most important ones in assisting the specialist. Thus, this work uses the ResNet architecture family and weighted metrics to evaluate hard samples on a breast ultrasound dataset that provides annotations from where a senior ultrasonographer evaluated the exam as benign or probably benign, but biopsy revealed a malignant result. As expected, these weighted metrics were mainly lower than traditional metrics, helping reduce the optimistic vision of model performance. ResNet 50 was the most stable model, getting at least one hard sample right in each fold.