Dr Fiona Maclean 1,2
1 Faculty of Medicine, Macquarie University, Macquarie Park, NSW, Sydney, Australia
2 Franklin.ai, Sydney, NSW, Australia
Oral presentation (link) at 36th European Congress of Pathology, 7 - 11 September 2024, Florence, Italy
Prostate carcinoma has a significant worldwide burden, with an expected 85% increase in deaths by 2040. We developed an Artificial Intelligence (AI) model that detects 45 clinical findings despite technical artefacts. This study validates the standalone performance of the model.
An ensemble AI model was trained on > 70,000 H+E stained whole slide images (WSIs) to detect, classify, localise and quantify routine findings across needle core biopsy (NCB) and transurethralresection of prostate (TURP) specimens. WSIs containing a variety of technical factors were included. The efficacy of the AI model was validated against > 1,700 ground truthed WSIs from distinct patients.
The AI model demonstrated strong performance in the detection and classification of 45 malignant and benign clinical findings across both NCB and TURP WSIs in the presence and absence of technical artefacts. Over 71 % of test set slides contained technical artifacts including staining artifacts, air bubbles, and tissue factors including tears, folds, calcification and diathermy artefact, with no significant loss of model performance. Robust results were obtained for identification and segmentation of acinar adenocarcinoma including identification of Gleason patterns 3-5. The model accurately detected cancer mimics including partial atrophy, post-atrophic hyperplasia, adenosis, Cowper’s glands and basal cell hyperplasia.
The standalone results of the AI model for prostate NCB and TURP WSIs in the presence and absence of technical artefacts demonstrate reliable classification and localisation of findings. Additionally, the model's performance did not diminish in the presence of known cancer mimics. As the burden of pathological diagnosis of prostate cancer increases, there is a growing need to provide support to pathologists. Utilisation of this AI model may mitigate this burden and growing demands.
DOI: https://doi.org/10.1007/s00428-024-03880-y
The results presented in this abstract are based on an early version of our AI model therefore outcomes reported here are subject to change as the model undergoes further development and validation.
Interested in learning more? Get in touch