Ovarian cancer affects women worldwide, and early and accurate detection is essential for appropriate treatment planning. The gold standard for diagnosis is histological biopsy, a costly and invasive procedure that could potentially be avoided in many cases through reliable imaging-based assessment. Ultrasound imaging represents a widely available and cost-effective alternative; however, its interpretation is inherently subjective and depends on the clinician’s expertise. In recent years, deep learning approaches have been applied to ultrasound images for automatic differential diagnosis, often achieving performance comparable to expert clinicians. Nevertheless there are some important limitations related to the application of standard supervised machine learning in this setting.