Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-task Learning Approach for Intracranial Hemorrhage Prognosis

Miriam Cobo, Amaia Pérez del Barrio, Pablo Menéndez Fernández-Miranda, Pablo Sanz Bellón, Lara Lloret Iglesias, Wilson Silva

TL;DR

This work tackles CT-based prognosis in intracranial hemorrhage by learning a shared representation that couples prognosis with key clinical variables. It introduces two end-to-end 3D multi-task models based on DenseNet121 that predict prognosis alongside discretized GCS and age, using a joint loss to regularize image features toward clinically meaningful information. The approach yields improved performance over image-only baselines and demonstrates competitive alignment with neuroradiologists on CT-only input, supported by interpretable saliency maps. By validating against a public dataset and providing code, the study advances interpretable multimodal prognosis and highlights the prognostic relevance of GCS and age in imaging models.

Abstract

Prognosis after intracranial hemorrhage (ICH) is influenced by a complex interplay between imaging and tabular data. Rapid and reliable prognosis are crucial for effective patient stratification and informed treatment decision-making. In this study, we aim to enhance image-based prognosis by learning a robust feature representation shared between prognosis and the clinical and demographic variables most highly correlated with it. Our approach mimics clinical decision-making by reinforcing the model to learn valuable prognostic data embedded in the image. We propose a 3D multi-task image model to predict prognosis, Glasgow Coma Scale and age, improving accuracy and interpretability. Our method outperforms current state-of-the-art baseline image models, and demonstrates superior performance in ICH prognosis compared to four board-certified neuroradiologists using only CT scans as input. We further validate our model with interpretability saliency maps. Code is available at https://github.com/MiriamCobo/MultitaskLearning_ICH_Prognosis.git.

Multi-task Learning Approach for Intracranial Hemorrhage Prognosis

TL;DR

This work tackles CT-based prognosis in intracranial hemorrhage by learning a shared representation that couples prognosis with key clinical variables. It introduces two end-to-end 3D multi-task models based on DenseNet121 that predict prognosis alongside discretized GCS and age, using a joint loss to regularize image features toward clinically meaningful information. The approach yields improved performance over image-only baselines and demonstrates competitive alignment with neuroradiologists on CT-only input, supported by interpretable saliency maps. By validating against a public dataset and providing code, the study advances interpretable multimodal prognosis and highlights the prognostic relevance of GCS and age in imaging models.

Abstract

Prognosis after intracranial hemorrhage (ICH) is influenced by a complex interplay between imaging and tabular data. Rapid and reliable prognosis are crucial for effective patient stratification and informed treatment decision-making. In this study, we aim to enhance image-based prognosis by learning a robust feature representation shared between prognosis and the clinical and demographic variables most highly correlated with it. Our approach mimics clinical decision-making by reinforcing the model to learn valuable prognostic data embedded in the image. We propose a 3D multi-task image model to predict prognosis, Glasgow Coma Scale and age, improving accuracy and interpretability. Our method outperforms current state-of-the-art baseline image models, and demonstrates superior performance in ICH prognosis compared to four board-certified neuroradiologists using only CT scans as input. We further validate our model with interpretability saliency maps. Code is available at https://github.com/MiriamCobo/MultitaskLearning_ICH_Prognosis.git.
Paper Structure (14 sections, 1 equation, 6 figures, 5 tables)

This paper contains 14 sections, 1 equation, 6 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Materials and Methods
  3. Data
  4. Method
  5. Preprocessing:
  6. Tabular models:
  7. Image models:
  8. Evaluation
  9. Results
  10. Tabular models
  11. Image models
  12. Discussion and Conclusion
  13. Acknowledgments.
  14. Disclosure of Interests.

Figures (6)

  • Figure 1: Proposed multi-task image model integrating GCS and age as outputs to regularize the learning and enhance the prognosis task. In the saliency maps, brighter colors mean higher importance.
  • Figure 2: Ablation analysis of each component of the loss in our method.
  • Figure 3: Guided-Backpropagation saliency maps ($p$ is the output probability, $p < 0.5$ corresponds to good prognosis). A: Good prognosis. Correctly labelled by 4/4 neuroradiologists. B: Good prognosis. Incorrectly labelled by 4/4 neuroradiologists. C: Poor prognosis. Incorrectly labelled by 4/4 neuroradiologists. D: Poor prognosis. Correctly labelled by 3/4 neuroradiologists.
  • Figure 4: Tabular models to identify the most relevant variables driving ICH prognosis predictions, reproduced with logistic regressor and decision tree classifier models trained on the two main variables guiding the decisions: GCS and age.
  • Figure 5: Guided-Grad-Cam saliency maps for four example test images ($p$ indicates the output probability of each model). Patient ID is indicated in brackets. A: Good prognosis (34). Correctly labelled by 4/4 neuroradiologists. B: Good prognosis (93). Incorrectly labelled by 4/4 neuroradiologists. C: Poor prognosis (59). Incorrectly labelled by 4/4 neuroradiologists. D: Poor prognosis (140). Correctly labelled by 3/4 neuroradiologists.
  • ...and 1 more figures