Soft multiclass feature augmented deep learning to predict tumor origins using cytology or histology whole slide images

Tailoring cancer treatment to the primary tumor site is essential for optimal outcomes, yet identifying the source of metastasis remains a significant clinical challenge. The primary tumor location influences not only the choice of treatment but also its delivery method, potential side effects and overall prognosis. Cancer of unknown primary (CUP) persists despite comprehensive diagnostic evaluations. We present the first fast, low-cost deep learning based cancer screening approach, i.e. a new data efficient soft multiclass feature augmentation deep learning, to identify the primary origin of malignant cells directly from whole slide images (WSIs) of cytological smears (CS), cell blocks (CB) or histopathology slides. Our framework was evaluated across three datasets: a multi-center histopathology collection of 69 sites (1,196 WSIs), a CS dataset (260 WSIs), and a CB dataset (129 WSIs). The proposed method achieved excellent accuracy, outperforming four state-of-the-art methods in identifying common metastatic origins. These findings demonstrate the potential of this deep learning tool to advance precision oncology by rapidly guiding clinicians toward appropriate, site-specific cancer treatments.