Fig. 3: Results for in-domain tasks.
From: Overcoming data scarcity in biomedical imaging with a foundational multi-task model
a, In the diagnosis of pneumonia (Pneumo-CXR), UMedPT matched the full fine-tuned performance of ImageNet, even with a frozen encoder and a reduced dataset size (1%). b, CRC-WSI was the only target task where the training dataset was also part of the pretraining. Here, performance was stable across dataset fractions with a frozen encoder. When the encoder was fine-tuned, performance decreased to the result obtained with ImageNet pretraining. c, For NucleiDet-WSI, an object detection task for counting nuclei in WSIs, UMedPT outperformed ImageNet across all training settings. Best performance was achieved with 100% of the training data and fine-tuning. In each setting, five independent trainings were derived for each training subset and method. The middle line of the boxes represents the median, the boundaries are the Q1 and Q3 quartiles, the whiskers extend to 1.5 times the interquartile range (IQR), and outliers beyond are shown as single points.
