ESTRO 2024 - Abstract Book

S4529

Physics - Machine learning models and clinical applications

ESTRO 2024

GBM dataset [5]. Only morphological sequences (T2, FLAIR, T1 and T1 post-contrast) were used for model development and testing.

Following the evaluation approach described in [6] for comparability, we developed two distinct classifiers for progression status prediction. The first classifier was designed to discern between PsP (with scores 1-2) and all other cases (scores 3-6), while the second classifier aimed to distinguish TP (with scores 5-6) from all other cases (scores 1 4).

Results:

In the TP versus all cases classification, our model achieved an accuracy of 90% and an AUC of 0.803. It also demonstrated a sensitivity of 66.6% and a specificity of 94.1%. Similarly, in the classification of PsP versus all other cases, our model showed promising results with an accuracy of 85% and an AUC of 0.833. In this context, the model exhibited a sensitivity of 100% and a specificity of 66.6% (Table 1). The findings underscore the effectiveness of the self-supervised learning approach in accurately distinguishing between TP and PsP with only a limited amount of labeled training data (n = 40).

Experiment

Accuracy

AUC

Sensitivity

Specificity

TP vs All

0.90

0.80

0.66

0.94

PsP vs All

0.85

0.83

1.00

0.66

3 Categories

0.70

-

-

-

Conclusion:

While achieving results that are in close proximity to the state-of-the-art [6], our method stands out for its reliance solely on morphological MRI sequences [T2, FLAIR, and T1 pre- and post-contrast], avoiding the need for advanced diffusion und perfusion sequences that have limited availability. This study advances the challenging task of distinguishing between PsP and TP in HGG MR imaging, a crucial aspect of timely clinical decision-making in radiation oncology. The self-supervised deep learning approach, trained on rich unlabeled MRI data, effectively overcomes data constraints and overfitting issues. Notably, it achieved a high accuracy in classifying TP (accuracy 90%, AUC 0.803) and a high sensitivity in distinguishing PsP (sensitivity 100%, AUC 0.833) (Figure 1). These results are competitive with the state-of-the-art, while our model's exclusive use of widely available MRI modalities enhances its practicality for clinical applications. The presented approach could have real-world clinical relevance, including healthcare settings with limited resources.