(UroToday.com) The 2023 American Society for Radiation Oncology (ASTRO) 65th Annual Meeting held in San Diego, CA between October 1st and 4th, 2023 was host to a session on radiotherapy and the post-prostatectomy setting. Dr. Atallah Baydoun presented the results of an analysis evaluating post-prostatectomy risk stratification in the biochemical recurrence setting using transfer learning-based multi-modal artificial intelligence.
Dr. Baydoun began by highlighting that for patients undergoing radical prostatectomy for prostate cancer, accurate risk stratification is essential to guide post-prostatectomy therapeutic decision making. Over the last two decades, multiple prognostic biomarkers have been explored in an effort to stratify biochemical risk in post-prostatectomy patients.
What are the pitfalls of the currently employed models?
- CAPRA-S: A ‘post-prostatectomy’ biomarker that depends on pathologic data from the radical prostatectomy specimen
- Gleason score: Scoring is heavily affected by inter-observe variability, which is most prominent when the pathology specimen is obtained from diagnostic biopsies (kappa=0.36), rather than surgical prostatectomy (kappa=0.45)
- Other prognostic biomarkers have limited prognostic abilities with AUCs <0.75, as summarized below:
Given these limitations, the authors sought to train and test a multi-modal artificial intelligence model (TRAIL: Transfer leaning-based multi-modal Artificial Intelligence model) for biochemical risk stratification following radical prostatectomy. To this end, the authors included patients who underwent a prostatectomy with data available within a single institution prospective registry. The unannotated digital pathology slides from the diagnostic biopsies prior to radical prostatectomy for patients with clinically localized disease were scanned at 20x resolution and included in this analysis.
To label their tissue, they first downloaded the publicly available PANDA dataset, which consists of approximately 11,000 whole-slide images of digitized H&E-stained biopsies originating from two centers. Next, the authors used a pre-trained inceptionv3 network. Thirdly, the authors adapted Inceptionv3 to the task of prostate malignant tissue labeling and training on the PANDA dataset, such that it outputs for each biopsy block sized pixel, a probability of the block having malignant prostate tissue, leading to a trained neural network, as summarized in the schematic below:
Next, using the PANDA challenge data trained Inceptionv3 network, heatmaps for cancerous blocks were generated.
After the heatmaps were generated with Inceptionv3, the areas with high risk of malignant tissue were extracted. Subsequently, these areas were divided into blocks sized [256 256 3] that served as input for the pre-trained AlexNet to derive AI-derived quantitative features.
In addition to digitized pathology slides of the diagnostic biopsy specimens, the following clinicopathologic data were available:
Finally, the AI and clinicopathologic variables were combined using RUSBoost to create TRAIL, which is a classification ensemble model that predicts two-year biochemical recurrence.
The total number of patients included in the TRAIL workflow was 376. The median patient age was 63 years (range: 41 – 75), median PSA was 6.5 ng/ml (range: 0.7 – 97), and the median follow-up was 53.5 months (range: 1 – 136). Of note, 98% of patients had pathologic Gleason score 8 disease, 38% had pT3 or worse disease, and 26% had positive margin status.
In the overall cohort, the biochemical recurrence rates were 12% at 2 years and 18% at 3 years.
Inceptionv3 achieved good accuracy for the delineation of cancerous areas on the diagnostic biopsy slides, with an accuracy of approximately 82% in the testing dataset.
The TRAIL model was trained on 257 patients, and then subsequently locked and applied to the test set of 119 patients. This model incorporated age, PSA, and fraction of positive biopsy cores. TRAIL achieved an accuracy of 0.95 in the testing set, with an area under the curve (AUC) of 0.81 for the prediction of biochemical recurrence, which outperforms other models, such as CAPRA-S (AUC: 0.57).
The prognostic ability of TRAIL is illustrated in the cumulative incidence function curves below, whereby we see that TRAIL positive patients had significantly increased incidence of biochemical recurrence, with an early separation of the curves at approximately 8 – 10 months.
After training and testing, the occlusion sensitivity algorithm was used to visually localize the areas within the slide driving the AI features classification. The visual feature areas were segmented using K-mean clustering. This demonstrated that the ‘hot spots’ were correlated with known adverse histopathologic features, such as clear vacuoles and vague lamina formation.
Dr. Baydoun concluded his presentation as follows:
- TRAIL is independent of prostatectomy data and can offer patients the advantage of decision-forming capacity at the time of pre-operative decision making, including the likelihood for post-operative treatment intensification
- The AI potential of constructing complex, high dimensional, non-linear models from multi-modal data is expanding the scope of prognostic biomarkers in medicine
- AI’s explainability for digital pathology remains at its early stages
- Future work with larger datasets with metastatic events is needed to further optimize the model for clinical use
Presented by: Atallah Baydoun, MD, PhD, Physician Resident, Department of Radiation Oncology, Case Western Reserve University, University Hospitals of Cleveland, Cleveland, OH
Written By: Rashid K. Sayyid, MD, MSc – Society of Urologic Oncology (SUO) Clinical Fellow at The University of Toronto, @rksayyid on Twitter during the 2023 American Society for Therapeutic Radiation Oncology (ASTRO) 65th Annual Meeting held in San Diego, CA between October 1st and 4th, 2023