EP175 - Improving rectal cancer care by developing an automated artificial intelligence algorithm for magnetic resonance imaging interpretation: a pilot study

Image or Table (Do Not add text here):

Purpose/Background: Pelvic magnetic resonance imaging (MRI) is the preferred method for characterising rectal cancer (RC) stage. Accurate evaluation of stage is critical for decision making. However, assessing pelvic MRI is challenging. Review by multiple interpreters can improve diagnostic precision, but discrepancies in clinical staging are well described in the literature. Improvement in MRI interpretation ought to improve patient care. Artificial intelligence (AI) has been applied successfully in the interpretation of a variety of imaging modalities, including RC MRI. However, incorporating an automated AI algorithm as an additional imaging interpreter in MTB has not yet been demonstrated, partly because AI typically needs experts to label large numbers of cases, a very time-consuming and costly task. We evaluated the accuracy level possible from an automated AI algorithm using only a tiny number of labelled images. This tool is intended for eventual incorporation in RC MTB boards to mitigate discrepancies, ease decision making, and ultimately, improve patient outcomes.

Methods/Interventions: A sample of 100 MRI images (1.5T, T2-weighted), most visibly containing tumour, were selected from 36 patients who underwent rectal cancer surgery from Nov 2018 - Feb 2023 at a single Canadian tertiary centre. These images were labelled for T-stage by an expert user, using 3D Slicer. In our initial testing, conventional segmentation models like UNet and UNet++ showed poor generalisation to unseen test images. To address this, we utilised extensive data augmentation and a pre-trained foundation model (Meta SAM2). We split the data, using 80% for training and 20% for testing, ensuring no patient overlap to avoid data leakage. We used a foundation model trained on 50.9k videos and 35.5M masks and fine-tuned it on our dataset. Our data augmentation to increase the dataset size and reduce overfitting included randomly applying horizontal/vertical flips, random rotations within a range of -90 to +90 degrees, and centre cropping to a scale range between 50% and 100% of the original image size. At each training step, we selected up to two of these transformations at random. Test results were examined by another expert and a qualitative score was awarded (grade 1 - incorrect T-staging, grade 2 - correct T-staging with incorrect contours, grade 3 - correct T-staging with correct contours). Finally, a Dice score was calculated that evaluated the extent of overlap between AI generated vs manually delineated tumour boundaries.

Results/Outcomes: On the 20 test images, our model achieved an average Dice score of 58.2%. A representative example of model results is seen in Figure 1, where ground-truth is labelled by the expert observer and the prediction is modelled by the AI algorithm. Dice score evaluates the precise lesion boundaries and may not accurately characterise network performance on clinical tasks. On our expert qualitative assessment, 65% of images had correct T-staging (grade 3, 35%, n=7, and grade 2, 30%, n=6).

Conclusion/Discussion: This pilot study represents the initial stages of the development of an automated AI algorithm for the interpretation of RC MRI. These Initial results show surprisingly promising AI performance despite use of only a tiny training set (just 80 images of 36 patients). This suggests it is likely feasible to train AI models with relatively small numbers of expert labels. Further multicentre training data will be acquired to improve the accuracy of the algorithm.