Slice Selection Guided by Grad-CAM in CT Scans for Brain Hemorrhage Detection

• Daniel C. Santos UFAM
• Paulo H. N. Gonçalves UFAM
• Eduardo Souto UFAM
• Robson L. O. Amorim UFAM

DOI: https://doi.org/10.5753/sbcas.2025.7752

Abstract

Intracranial hemorrhages are severe and potentially fatal conditions that require rapid and accurate diagnosis to reduce neurological sequelae and mortality. Computed tomography (CT) is widely used for this detection, but manual evaluation is time-consuming and subject to interobserver variability. Three-dimensional convolutional neural networks (CNN3D) have been applied to assist in volumetric medical image analysis. However, their use is hindered by high computational cost and redundant information in full CT volumes. This study proposes Gradient-weighted Class Activation Mapping (Grad-CAM) 2D Guided Selection (SGG-2D), a hybrid approach that combines adaptive and non-adaptive techniques for efficient slice selection. The method employs a 2D classification model with Grad-CAM to identify the most informative slices, complemented by a structured sampling strategy that ensures a balanced distribution of selected slices across the tomographic volume. The approach was compared against six existing methods, achieving the highest accuracy (75.67%) and AUC-ROC (82.72%). The results demonstrate that optimized slice selection not only reduces computational complexity but also enhances the quality of data input for CNN3Ds, leading to improved accuracy in anomaly classification in volumetric exams.

References

Anouk Stein, M., Wu, C., Carr, C., Shih, G., Kalpathy-Cramer, J., Elliott, J., kalpathy, Prevedello, L., Marc Kohli, M., Lungren, M., Culliton, P., Ball, R., and MD, S. H. (2019). Rsna intracranial hemorrhage detection. [link]. Kaggle.

Bai, Q., Gao, Y., Chen, F., Zhang, Y., Yang, Y., Zhong, L., Yang, W., and Yang, Y. (2024). Representative slice selection and multi-view projection learning for pulmonary tuberculosis infectiousness identification using ct volume. IEEE Access, 12:1374–1390.

Burduja, M., Ionescu, R. T., and Verga, N. (2020). Accurate and efficient intracranial hemorrhage detection and subtype classification in 3d ct scans with convolutional and long short-term memory neural networks. Sensors, 20(19):5611.

da Silva, L. A. (2023). Abordagem de aprendizado profundo para extração de quadros significativos em volumes de tomografia computadorizada. Dissertação de mestrado, Universidade Federal do Amazonas - ICOMP-UFAM, Manaus, Brasil.

Köpüklü, O., Kose, N., Gunduz, A., and Rigoll, G. (2019). Resource efficient 3d convolutional neural networks. arXiv preprint arXiv:1904.02422.

Oladimeji, O., Ayaz, H., Unnikrishnan, S., and McLoughlin, I. (2023). Lightweight deep learning for breast cancer diagnosis based on slice selection techniques. In Proceedings of the 2023 Irish Conference on Artificial Intelligence and Cognitive Science (AICS). IEEE.

Xie, S., Girshick, R., Dollár, P., Tu, Z., and He, K. (2017). Aggregated residual transformations for deep neural networks. arXiv preprint arXiv:1611.05431.

Zunair, H., Rahman, A., Mohammed, N., and Cohen, J. P. (2019). Estimating severity from ct scans of tuberculosis patients using 3d convolutional nets and slice selection. CLEF2019 Working Notes.

Zunair, H., Rahman, A., Mohammed, N., and Cohen, J. P. (2020). Uniformizing techniques to process ct scans with 3d cnns for tuberculosis prediction. In Maglogiannis, I., Iliadis, L., and Pimenidis, E., editors, Predictive Intelligence in Medicine, pages 156–168. Springer International Publishing.