Autism Detection based on Structural MRI using YOLO

Technical Summary: Autism Detection based on Structural MRI using YOLO

Problem Statement
Autism Spectrum Disorder (ASD) is a neurodevelopmental condition affecting social communication and behavior, with prevalence rates rising significantly (e.g., 1 in 36 children in 2020). Early diagnosis is critical for intervention but is often delayed due to reliance on subjective behavioral assessments and a lack of objective biological markers. While Structural Magnetic Resonance Imaging (sMRI) offers a non-invasive method to detect neuroanatomical abnormalities associated with ASD, existing deep learning approaches face several limitations. Many studies utilize broad age ranges without accounting for age-related brain developmental differences, leading to poor generalization and overfitting. Furthermore, traditional classification models (e.g., ResNet50, VGG16) often treat MRI scans as global image classes, potentially overlooking subtle, localized structural biomarkers. There is a need for efficient, intelligent screening tools that leverage advanced object detection architectures to identify these subtle neuroanatomical deviations with high accuracy.

Methodology
The study proposes a diagnostic framework utilizing You Only Look Once (YOLO) object detection models (versions 8 through 11) applied to sMRI data from the ABIDE II dataset. The methodology involves the following key steps:

1. Dataset and Preprocessing:

   • Data Source: The ABIDE II dataset, containing T1-weighted sMRI scans from 522 individuals with ASD and 592 typically developing (TD) individuals.
   • Age Stratification: To address developmental variability, the dataset was divided into two specific age groups: 5.8–6.8 years and 7–12 years.
   • Slice Extraction: Raw 3D sMRI volumes were converted into 2D image sequences. From each of the three anatomical planes (axial, sagittal, and coronal), the 50 central slices containing the most vital brain information were extracted.
   • Annotation: Using the Roboflow platform, each 2D slice was resized to 224 × 224 pixels. Unlike traditional segmentation, a single bounding box was drawn around the entire brain tissue area for each slice. The classification label (ASD or TD) was assigned to the entire slice based on the patient's diagnosis, treating the brain region as a single "object" to be detected.

2. Model Architecture:

   • The study adapted YOLO models (v8, v9, v10, v11), originally designed for object detection, for ASD classification. This approach leverages spatial location awareness and hierarchical feature networks (FPNs) to detect structural biomarkers as local patterns rather than global intensity patterns.
   • Key Architectural Features:
     • YOLOv8: Utilizes CSPNet backbones and FPN+PAN necks with anchorless detection.
     • YOLOv9: Employs Generalized Efficient Layer Aggregation Network (GELAN) and Programmable Gradient Information (PGI) for improved multi-scale feature fusion.
     • YOLOv10: Uses CSPNet with depth-separable convolutions for reduced computational cost.
     • YOLOv11: Incorporates C2PSA (Cross-Stage Partial with Parallel Spatial Attention) mechanisms to focus on specific regions of interest and SPPF for spatial feature aggregation.
   • Training Configuration: Models were trained on Google Colab (NVIDIA T4) using PyTorch. Parameters included a batch size of 16, 100 epochs, and specific learning rates (0.001 for Adam). Data augmentation and early stopping (patience = 20) were employed to prevent overfitting.

3. Evaluation Metrics:

   • Performance was assessed using Precision, Recall, Mean Average Precision (mAP50 and mAP50-95), and the Average Dice Similarity Coefficient (Av DICE) to measure spatial overlap between predictions and ground truth.
   • Comparative baselines included traditional CNN architectures: ResNet50 and VGG16.

Key Contributions

• Application of Object Detection to ASD: The study demonstrates the efficacy of adapting YOLO object detection models for neuroimaging classification, framing the detection of ASD biomarkers as an object detection task rather than a standard image classification problem.
• Age-Stratified Analysis: By explicitly separating the dataset into two distinct younger age groups (5.8–6.8 and 7–12 years), the study addresses the limitation of broad age-range studies, aiming to improve diagnostic generalization across developmental stages.
• Multi-Plane Slice Extraction: The framework generates 2D image sequences from 3D sMRI data by extracting 50 slices from axial, sagittal, and coronal planes, maximizing the spatial information available for training.
• Comparative Performance: The research provides a comprehensive comparison of YOLO versions 8–11 against traditional deep learning models (ResNet50, VGG16) specifically for ASD detection on sMRI data.

Experimental Results
The experiments yielded the following findings:

• Superiority of YOLO Models: Across both age groups, YOLO models consistently outperformed ResNet50 and VGG16 in precision, recall, mAP, and DICE scores.
• Age Group 1 (5.8–6.8 years): YOLOv11 achieved the highest performance with a precision of 0.9957, recall of 0.998, and an average DICE coefficient of 0.9978. YOLOv9 also performed exceptionally well. In contrast, ResNet50 and VGG16 showed lower accuracy and signs of overfitting or poor generalization.
• Age Group 2 (7–12 years): All YOLO versions maintained high performance (precision and recall generally ≥0.95). YOLOv9 and YOLOv8 achieved the highest DICE scores (0.9957), while YOLOv11 showed a slight relative decrease in performance compared to the younger group, though still outperforming traditional CNNs.
• Model Robustness: The results suggest that YOLO architectures, particularly those with advanced attention mechanisms (like C2PSA in v11) and efficient feature aggregation (like GELAN in v9), are better suited for identifying the subtle, localized structural abnormalities in sMRI associated with ASD.

Significance and Claims
The paper claims that the proposed framework offers a promising insight into the early diagnosis of ASD using YOLO technology. The study highlights that:

• Object detection architectures can be successfully extended to neuroimaging classification by treating structural differences as local patterns.
• The integration of multiscale features and spatial attention processes in YOLO models (specifically v9 and v11) effectively reveals subtle neuroanatomical abnormalities in the early stages of ASD.
• The framework demonstrates high computational efficiency and robustness in handling MRI images.
• While the results are outstanding, the authors position this work as a step toward future clinical applicability, suggesting that such technology could support early detection and personalized intervention strategies. The study does not claim immediate clinical deployment but rather validates the technical feasibility and superior performance of this approach compared to existing methods.