CHIASM-Net: Artificial Intelligence-Based Direct Identification of Chiasmal Abnormalities in Albinism

Abstract

Purpose: Albinism is a congenital disorder affecting pigmentation levels, structure, and function of the visual system. The identification of anatomical changes typical for people with albinism (PWA), such as optic chiasm malformations, could become an important component of diagnostics. Here, we tested an application of convolutional neural networks (CNNs) for this purpose.

Methods: We established and evaluated a CNN, referred to as CHIASM-Net, for the detection of chiasmal malformations from anatomic magnetic resonance (MR) images of the brain. CHIASM-Net, composed of encoding and classification modules, was developed using MR images of controls (n = 1708) and PWA (n = 32). Evaluation involved 8-fold cross validation involving accuracy, precision, recall, and F1-score metrics and was performed on a subset of controls and PWA samples excluded from the training. In addition to quantitative metrics, we used Explainable AI (XAI) methods that granted insights into factors driving the predictions of CHIASM-Net.

Results: The results for the scenario indicated an accuracy of 85 ± 14%, precision of 90 ± 14% and recall of 81 ± 18%. XAI methods revealed that the predictions of CHIASM-Net are driven by optic-chiasm white matter and by the optic tracts.

Conclusions: CHIASM-Net was demonstrated to use relevant regions of the optic chiasm for albinism detection from magnetic resonance imaging (MRI) brain anatomies. This indicates the strong potential of CNN-based approaches for visual pathway analysis and ultimately diagnostics.

Figure 1.

Differences between averaged images of controls and albinism from the TRAIN dataset from the fold 5. (A) Slices from averaged images of controls. (B) Binary mask of voxels (intensity threshold: <0.75) from the averaged control image. (C) Difference between averaged control and albinism images masked by the binarized image B. (D) Difference between averaged control and albinism images. (E) Difference between averaged albinism and controls images masked by the binarized image F (note the inverted intensity as compared to columns C and D). (F) Binary mask of voxels (intensity threshold: <0.6) from the averaged albinism image. (G) Averaged albinism images. All presented images use fixed heatmap with constant range of values. Images for albinism demonstrate higher intensity in the central parts of chiasm (blue arrows) as compared to healthy controls (column E), which indicates a more compact structure. In contrast, controls show higher intensities on the lateral parts of optic tracts and the chiasm (red arrows), which indicates an increased width of the latter (column C).

Figure 2.

Inspection of predicting process of CHIASM-Net using image of averaged controls (A) and albinism (B). The first column demonstrates subsequent averaged slices of control (A) and albinism (B), from inferior to superior slice. The following three columns indicate the regions of input driving the predictions, as assessed with, from left to right, saliency maps, DeepLIFT, and Occlusion methods. For “Saliency maps” and “DeepLIFT” hot colors mark areas where high voxel intensities positively contribute to the prediction of albinism. In the case of “Occlusion,” warmer colors indicate areas that contribute strongly to the outcome predictions. It should be noted that panels (A) and (B) use different value ranges. All XAI methods consistently indicate that higher intensities of central voxels of the optic chiasm positively contribute to the prediction of albinism (red arrows), whereas the higher intensity in the lateral parts of the optic chiasm has the opposite effect (blue arrows).

Figure 3.

Explanation of predictions for selected control and PWA input with occlusion method (fold 6, TEST1). (A) Sample control image. (B) Sample albinism image. Each panel in the left column demonstrates subsequent slices of T1w image, from inferior to superior. The right hand-column presents the attributions map obtained with occlusion approach. Warmer colors mark areas that mostly contribute to predictions. The predictions are driven by the areas containing the white matter which correspond to the optic chiasm and optic tracts.