Explainability and Clinical Trust of Deep Learning in Glioblastoma Treatment Efficacy Prediction: A Comprehensive Framework.
Authors
Affiliations (9)
Affiliations (9)
- Medical Physics Department, Centre François Baclesse, Caen, 14000, France.
- Artificial Intelligence Department, Centre François Baclesse, Caen, 14000, France.
- Université de Caen Normandie, CNRS, Normandie Univ, ISTCT UMR6030, CYCERON, Caen, 14000, France.
- Radiation Oncology Department, Centre François Baclesse, Caen, 14000, France.
- Radiology Department, Centre François Baclesse, Caen, 14000, France.
- Université Caen Normandie, ENSICAEN, CNRS, Normandie Univ, GREYC UMR 6072, Caen, 14000, France.
- Medical Physics Department, Centre François Baclesse, Caen, 14000, France. [email protected].
- Artificial Intelligence Department, Centre François Baclesse, Caen, 14000, France. [email protected].
- Université de Caen Normandie, CNRS, Normandie Univ, ISTCT UMR6030, CYCERON, Caen, 14000, France. [email protected].
Abstract
Glioblastoma (GBM) treatment response is highly heterogeneous. Predicting outcomes from pre-treatment MRI is a critical unmet need, yet the clinical adoption of "black box" deep learning (DL) models is hindered by a lack of transparency. Robust and holistic validation frameworks still need to be developed to allow clinical trust. A multi-layered eXplainable AI (XAI) pipeline was employed to dissect a previously developed deep learning model trained to predict GBM overall survival. First, the model's visual focus was qualitatively and quantitatively analysed. Second, the model's feature dependence was quantified using input ablation studies. Third, the depth at which class information becomes decodable was traced by linear probing of the internal representations. Finally, the clinical impact of the AI and its explanations was assessed in a multi-reader, multi-case study (37 patients, 4 readers, 3 phases, 444 observations) using Generalised Estimating Equations (GEE) and clustered bootstrapping. Analysis revealed a high-intensity heuristic: hyper-intense pixels at the skull periphery anchored "Long-Term" predictions, while the model defaulted to "Short-Term" based on tumour enhancement in its absence. The ablation study established this dependence, and linear probing pointed in the same direction. In the clinical study, AI assistance significantly improved expert accuracy (OR 1.598, p < 0.01) without artificially boosting confidence (p = 0.561). Importantly, the addition of XAI heatmaps was associated with a change in how experts used the AI score (interaction p < 0.01), and with fewer patients wrongly called long-term survivors. High predictive accuracy can mask reliance on low-level intensity heuristics. While superficial interpretation of saliency maps is perilous, multi-layered audits expose these shortcuts. Clinically, spatial heatmaps were associated with improved expert diagnostic accuracy and a shift in decision-making strategy.