An explainable three dimension framework to uncover learning patterns: A unified look in variable sulci recognition

TL;DR

This paper tackles the challenge of producing reliable global explanations for 3D neuroimaging models, focusing on the paracingulate sulcus (PCS) classification in 596 TOP-OSLO MRIs. It introduces an XAI 3D-Framework that fuses GradCam and SHAP extended to 3D with Shape-based dimensionality reduction, yielding robust global explanations via PCA across six components and optimized weighting. The authors demonstrate that the framework achieves superior faithfulness compared to single-method explanations and identifies key sub-regions—such as the posterior temporal and internal parietal areas, as well as cingulate and thalamic regions—correlated with PCS presence or absence. The approach advances interpretability in neuroimaging, enabling broader insights into developmental and pathological brain patterns, and lays the groundwork for applying 3D global explainability to other cortical morphology tasks and cohorts.

Abstract

The significant features identified in a representative subset of the dataset during the learning process of an artificial intelligence model are referred to as a 'global' explanation. 3D global explanations are crucial in neuroimaging, where a complex representational space demands more than basic 2D interpretations. However, current studies in the literature often lack the accuracy, comprehensibility, and 3D global explanations needed in neuroimaging and beyond. To address this gap, we developed an explainable artificial intelligence (XAI) 3D-Framework capable of providing accurate, low-complexity global explanations. We evaluated the framework using various 3D deep learning models trained on a well-annotated cohort of 596 structural MRIs. The binary classification task focused on detecting the presence or absence of the paracingulate sulcus, a highly variable brain structure associated with psychosis. Our framework integrates statistical features (Shape) and XAI methods (GradCam and SHAP) with dimensionality reduction, ensuring that explanations reflect both model learning and cohort-specific variability. By combining Shape, GradCam, and SHAP, our framework reduces inter-method variability, enhancing the faithfulness and reliability of global explanations. These robust explanations facilitated the identification of critical sub-regions, including the posterior temporal and internal parietal regions, as well as the cingulate region and thalamus, suggesting potential genetic or developmental influences. Our XAI 3D-Framework leverages global explanations to uncover the broader developmental context of specific cortical features. This approach advances the fields of deep learning and neuroscience by offering insights into normative brain development and atypical trajectories linked to mental illness, paving the way for more reliable and interpretable AI applications in neuroimaging.

Figures (8)

• Figure 1: The input modalities and the architecture of simple-3D-CNN and the simple-3D-MHL networks. a, Input modalities illustrated on a right hemisphere coronal slice. In more detail, the raw MRI of a given subject, the corresponding grey-white surface, and the corresponding sulcal skeleton. b, The architecture of simple-3D-CNN network with explanation of 3D Convolution layer (3D Conv) and 3D Max Pooling layer on the left. c, The three dimension MHL model with two different backbone choices, the full simple-3D-CNN (simple-3D-MHL) and the two level simple-3D-CNN layer (2CNN-3D-MHL).
• Figure 2: Classification task determination and the transition from local to global 3D explaination. a, Illustration of the no PCS condition (no PCS), and the PCS condition (PCS, in green line) on two left hemisphere 3D white matter reconstructions obtained with BrainVISA. The cingulate sulcus is coloured in yellow and blue and the callosal sulcus is coloured in purple. b, The 3D explainable framework that provides both local and global interpretations and explanations of our deep learning 3D classification network's results. The ratio of the faithfulness and complexity metrics were computed at that stage. In this example we include only the GradCam explainability method for simplicity.
• Figure 3: The proposed global 3D-Framework explanation. A weighted averaging (Weight tensor: [0.85, 0.7, 0.5, 0.3, 0.1, 0.001]) of six PCA components produces the average PCA image for PCA-Shape, PCA-GradCam, and PCA-SHAP. Following this, a weighted averaging (Weight tensor: [0.85, 0.5, 0.1]) of the three Global PCA overlapping images extracted the total overlapping image. This total overlapping image was then registered on a sulcal probabilistic atlas (the ICBM 2009a Nonlinear Asymmetric atlas, atlas1atlas2) to unveil the model's pattern for determining the presence or absence of the PCS.
• Figure 4: Simple-3D-MHL results on the left and right hemisphere of grey-white surface brain inputs. a,b, show the explainability results for the PCS class images of the first component among the six components of PCA for the total input modality (PCA-Shape), the total corresponding GradCam results (PCA-GradCam), and the total corresponding SHAP results (PCA-SHAP). The feature's importance (pixel attribution) varies from 0 (blue color) to 1 (red color), with high importance being 1 for the PCA-GradCam and PCA-Shape results. The orientation of the results are based on the medial anatomical views. All the presented results are align and mapping in the ICBM 2009a Nonlinear Asymmetric atlas (atlas1atlas2).
• Figure 5: Simple-3D-MHL results on the left and right hemisphere of sulcal skeleton brain inputs. a-b, show the explainability results for the noPCS class images of the first component among the six components of PCA for the total input modality (PCA-Shape), the total corresponding GradCam (PCA-GradCam), and the total corresponding SHAP results (PCA-SHAP). The feature's importance (pixel attribution) varies from 0 (blue color) to 1 (red color), with high importance being 1 for the PCA-GradCam and PCA-Shape results. The orientation of the results are based on the medial anatomical views. All the presented results are align and mapping in the ICBM 2009a Nonlinear Asymmetric atlas (atlas1atlas2).