Abstract: Accurate assessment of disease severity is often required for plant phenotyping, resistance breeding, and genetic analysis. In foxtail millet blast, severity is typically quantified as the proportion of lesion area relative to the leaf area. A critical indicator can be used to evaluate host resistance and downstream genetic studies. However, conventional visual scoring methods can rely heavily on human observation and experience, resulting in strong subjectivity, limited reproducibility, and coarse discrete grading. There is a high demand to capture continuous variations in disease progression. Fortunately, deep learning has significantly improved the performance of plant disease analysis in recent years. Most existing approaches have focused on image-level classification or pixel-level semantic segmentation. A systematic framework is lacking to convert segmentation outputs into reliable and continuous leaf-level severity measurements. Lesion detection and localization can be attributed to treating segmentation as the final optimization. Moreover, segmentation models often suffer from structural inconsistency and prediction fragmentation under complex lesion patterns, leading to substantial errors in area quantification at acceptable pixel-level metrics. In this study, a segmentation-driven quantification framework was developed to estimate continuous lesion severity in millet blast disease. Leaf and lesion regions were simultaneously predicted at the pixel level. Subsequently, a systematic computation was conducted on the ratio between lesion area and effective leaf area to derive a continuous severity indicator. Since accurate area estimation was sensitive to geometric distortions, high-resolution leaf images were acquired using a scanning-based strategy, rather than conventional camera-based imaging. Artifacts, such as leaf curling and perspective deformation, provided a more stable basis for quantitative analysis. Segmentation performance was further enhanced under limited data and high-resolution inputs. An adapter fine-tuning strategy was adopted on a lightweight segment anything model variant (mobile segment anything model). Specifically, a progressive interaction feature adapter (PIFA) was introduced into the side-branch structure to progressively interact with multi-scale features from the backbone, enabling effective fusion of local and global representations for the pretrained knowledge. In addition, an adaptive frequency dynamic filtering module (AFDF) was designed for frequency-domain responses through input-conditioned dynamic filters, allowing the network to better capture irregular lesion structures and fine-grained boundaries. Furthermore, a deformable spatial attention module (DSA) was incorporated to enhance spatial continuity and directional feature aggregation, which was particularly beneficial to suppress fragmented predictions for the integrity of the regions. Experiments demonstrated that the proposed method consistently outperformed representative segmentation models in lesion segmentation tasks. Specifically, LQNet achieved a lesion IoU of 74.64% and an F1-score of 85.48%, outperforming all compared methods and improving upon the second-best SAM-Adapter by 1.29 and 0.86 percentage points, respectively. Superior Intersection over Union and F1-score were achieved in the challenging scenarios characterized by dense, small-scale, and highly fragmented lesions. In terms of lesion severity estimation, LQNet achieved an RMSE of 1.067 and an R² of 0.954, with most prediction errors within ±1%, indicating improved accuracy and robustness in quantitative analysis. In discrete severity grading, the macro-average accuracy and micro-average accuracy reached 97.92% and 97.06%, respectively. Prediction fragmentation was reduced to bridge the gap between pixel-level segmentation and leaf-level quantification. In addition, instance-level analysis further revealed that the false positive detections were reduced in the mismatched fragmented regions, thereby improving precision and overall reliability. Overall, the framework can provide an effective solution to transform segmentation outputs into stable continuous severity indicators, indicating strong potential for practical applications in quantitative plant disease analysis, resistance evaluation, and genetic research.