Session Item

The purpose of this work was to assess whether archived prostate core biopsy can be used to predict long-term disease recurrence using artificial Intelligence driven digital pathology.

Cancer is a disease of the genome involving genetic and epigenetic alterations of the genetic material found within the cell’s nucleus. Using a DNA specific stain (Feulgen-Thionin) we performed quantitative measurements of the amount, distribution and spatial organization of the DNA within individual prostate cell nuclei in tissue sections from needle biopsies.  120 patients used for tis analysis had a spectrum of prostate cancer disease from low to high risk disease. They wer treated with LDR brachytherapy +- hormone therapy +- external beam radiation therapy. Median follow up was 100 months, (range 42-186mo). DNA measures were used to differentiate patients which had PSA biochemically identified recurrence within 36 months (60 sections from 19 cases), after 36 months (41 sections from 16 cases), or were recurrence free for more than 5 years (163 sections from 85 cases). 

Using AI methods for each section well segmented epithelial/tumour nuclei were identified and subdivided into nuclei presenting with characteristics associated with recurrence within 36 (poor outcome) months or characteristics associated with no recurrence after 5 years or more (good outcome). A Bad Cell Ratio (number of poor outcome characteristic nuclei / [the number of poor outcome characteristic nuclei + the number of good outcome characteristic nuclei]) was calculated per section and a Bad Cell Ratio greater than a threshold predicted poor outcome. On the 173 training set sections this resulted in a 94.2% correct prediction accuracy per section and a 100% accuracy at the patient level. On the 50 test set sections this threshold resulted in a 96.1% correct prediction accuracy per section and a 97% accuracy at the patient level. On the 41 validation set sections this process resulted in 88% correct prediction per section and a 94% accuracy at the patient level. 

Figure 1. Kaplan-Meier curves for the Training and Test sets. Results are reported per section

Figure 2. ROC for the Training and Test sets. Results are reported per section.

Using Artificial Intelligence assisted digital pathology,  and DNA specific stain,  to performed quantitative measurements of the amount, distribution and spatial organization of the DNA within individual prostate cell nuclei in tissue sections from prostate needle biopsies, we could predict the long term PSA recurrence in patients treated with brachytherapy +- without hormone  therapy or external beam radiation. Ultimately this information could be used to individualized the treatment intensity an achieve  optimal oncological outcomes.