Fig. 4: Results for out-of-domain tasks.
From: Overcoming data scarcity in biomedical imaging with a foundational multi-task model
a, For tuberculosis diagnosis (Tuber-CXR), UMedPT enables the full ImageNet performance to be matched, even with frozen encoder and a reduced amount of data (50%). b, In BC-Bach-WSI, fine-tuning was necessary for high performance. In the frozen setting, half of the training data were required to match the ImageNet frozen result. In each setting, five independent trainings were derived for each training subset and method. The middle lines of the boxes represent the median, the boundaries are the Q1 and Q3 quartiles, the whiskers extend to 1.5 IQR, and outliers beyond are shown as single points.
