ROI-Guided Lesion Segmentation for Maize Disease Phenotyping in Field Images

Technical Summary: ROI-Guided Lesion Segmentation for Maize Disease Phenotyping in Field Images

Problem Statement
Accurate delineation of disease lesions in field images is critical for objective crop health monitoring and phenotyping of maize foliar diseases. However, variable illumination, cluttered backgrounds, and irregular lesion morphologies (e.g., Northern Leaf Blight and Gray Leaf Spot) significantly reduce the reliability of lesion segmentation in real-world conditions. While object detection frameworks like YOLO can localize symptomatic regions, and semantic segmentation networks (e.g., DeepLabV3+) can delineate lesions, the specific impact of the spatial quality of detector-generated Regions of Interest (ROIs) on downstream segmentation performance remains unclear. Existing literature often modifies both detection and segmentation stages simultaneously or evaluates them in isolation, failing to determine whether a lightweight detector (under 3 M parameters) can generate ROIs with sufficient spatial fidelity to improve end-to-end segmentation without altering the segmentation architecture.

Methodology
The authors propose a controlled, two-stage detection-segmentation framework designed to isolate the effect of ROI quality on lesion delineation:

1. Detection Stage (DT-YOLO): A lightweight detector derived from YOLOv11n is developed to localize disease symptoms. To enhance efficiency and feature representation for irregular lesions, three specific modifications are integrated:

   • C3k2_DualPConv: Standard C3k2 modules are replaced with DualPConv blocks, utilizing Partial Convolution (PConv) to reduce computational redundancy while maintaining receptive field properties.
   • Post-Neck Triplet Attention: A lightweight Triplet Attention module is inserted after the neck (FPN-PAN) but before the detection head. This recalibrates dependencies across channel, height, and width dimensions without disrupting gradient flow in the fusion pathways.
   • MPDIoU Loss: The bounding box regression loss is switched from CIoU to Minimum Point Distance Intersection over Union (MPDIoU). This metric penalizes the Euclidean distance between homologous vertices of predicted and ground-truth boxes, better accommodating the irregular, non-elliptical geometries of maize lesions.

2. Segmentation Stage (DeepLabV3+): A standard DeepLabV3+ network with a ResNet50 backbone is used for pixel-level lesion mask generation. Crucially, this network remains architecturally unmodified throughout the study. It receives ROIs cropped from the original images based on the detector's predictions (with 10% padding to avoid truncating diffuse boundaries).

3. Experimental Design:

   • Dataset: 4,126 original maize leaf images (Healthy, GLS, NLB, CR, MSV) were augmented to 7,427 for detection. For segmentation, 600 images were manually annotated with pixel-level masks. An independent set of 100 images was used for end-to-end evaluation.
   • Evaluation: The study introduces an ROI Alignment Score to quantify the spatial agreement between detector-generated ROIs and ground-truth bounding boxes. The framework is compared against a baseline cascade (YOLOv11n + DeepLabV3+) and various single-stage instance segmentation models.

Key Contributions

1. Controlled Two-Stage Framework: A modular pipeline that isolates the downstream impact of detector-generated ROIs from segmentation network modifications, allowing for independent evaluation of localization and delineation.
2. DT-YOLO Detector: A lightweight model (2.52 M parameters, 6.40 GFLOPs) integrating C3k2_DualPConv, Triplet Attention, and MPDIoU loss, optimized for computational efficiency and geometric accuracy in field conditions.
3. ROI Alignment Score: A novel metric quantifying the spatial suitability of extracted ROIs for downstream segmentation tasks.
4. Empirical Analysis: A systematic investigation linking detector performance and ROI alignment to end-to-end segmentation results, highlighting the trade-offs between conventional detection metrics and downstream utility.

Results

• Detection Performance: DT-YOLO achieved 89.0% mAP@0.5, a 0.4 percentage point improvement over the YOLOv11n baseline (88.6%), with a slight reduction in parameters (2.52 M vs. 2.59 M) and FLOPs. However, it showed lower mAP@0.5:0.95 (53.9% vs. 55.9%), precision (82.9% vs. 86.1%), and recall (79.4% vs. 85.0%) compared to the baseline. The improvements were most pronounced for diseases with irregular morphologies (NLB and CR).
• ROI Alignment: The ROI Alignment Score increased significantly from 0.382 (YOLOv11n) to 0.635 (DT-YOLO), indicating a 66.2% relative improvement in spatial agreement.
• End-to-End Segmentation: When coupled with the unchanged DeepLabV3+, the DT-YOLO pipeline achieved 63.20% mIoU and 65.95% Dice, compared to 62.96% mIoU and 64.98% Dice for the YOLOv11n baseline.
• Comparison: While single-stage models (e.g., YOLOv11-Seg-N, Mask R-CNN) achieved higher raw segmentation accuracy, the proposed two-stage approach offers architectural modularity. The total computational cost of the two-stage pipeline (29.20 M parameters) is higher than lightweight single-stage detectors, but the design prioritizes the ability to update the localization component independently.

Significance and Claims
The paper claims that the study provides a modest but significant demonstration that the spatial quality of detector-generated ROIs directly influences downstream lesion segmentation, even when the segmentation network is held constant. The primary significance lies in the modular approach, which allows researchers and practitioners to independently evaluate and improve localization and segmentation components.

The authors emphasize that while the end-to-end performance gain is modest, the framework validates the hypothesis that a lightweight detector can produce ROIs with sufficient spatial fidelity to support effective pixel-level prediction. The study concludes that localization errors remain a major limitation for end-to-end performance, as evidenced by the gap between the end-to-end results and the "oracle" evaluation (where manually cropped regions were used). The framework is presented as a tool for maize disease phenotyping that aligns with crop protection decision-making by focusing computation on symptomatic tissue, though the authors caution that further validation on diverse field conditions and edge hardware is required before routine deployment.