Abstract: Accurate and continuous assessment of disease severity is essential for resistance breeding and genetic association analysis of foxtail millet blast, in which severity is commonly quantified as the ratio of lesion area to leaf area. However, traditional visual grading is subjective and coarsely discretized, and existing deep learning methods focus on lesion segmentation or classification without systematically converting segmentation outputs into leaf-level continuous severity indicators. This study therefore aimed to develop a segmentation-driven framework for objective, continuous, and repeatable severity quantification. A semantic segmentation-based quantification framework was proposed to simultaneously segment leaf and lesion regions at the pixel level and compute the lesion-to-leaf area ratio as a continuous severity indicator. Considering limited samples and high-resolution inputs, an adapter fine-tuning model, lesion quant net (LQNet), was constructed based on the lightweight mobile segment anything model (MobileSAM), in which a progressive interaction feature adapter (PIFA), an adaptive frequency dynamic filtering module (AFDFM), and deformable spatial attention (DSA) were designed to fuse multi-scale semantics, enhance complex lesion and boundary modeling, and improve spatial structural continuity, respectively. A high-resolution scanned leaf dataset was additionally built to reduce geometric errors caused by leaf deformation. Experiments were conducted on a self-constructed high-resolution scanned dataset containing 230 leaf instances and 25,204 lesion instances, with a 7:1.5:1.5 training/validation/test split. On the test set, LQNet achieved an intersection over union (IoU) of 74.64% and an F1-score of 85.48% for lesion segmentation, both the highest among UNet, DeepLabv3+, SwinUNet, DCSwin, and SAM-Adapter, with improvements of 1.29 and 0.86 percentage points over the strongest baseline SAM-Adapter. The leaf-region IoU and F1-score reached 95.46% and 97.67%, respectively, and the overall accuracy reached 98.46%, comparable to the highest value among the compared methods. Qualitative analysis further showed that LQNet produced more continuous and boundary-accurate predictions, especially for dense, small-scale, and highly fragmented lesions. Ablation experiments demonstrated that each module contributed stable gains and that their combination was the most effective, raising the lesion IoU by 4.94 and the F1-score by 3.34 percentage points over the baseline, verifying the synergy among PIFA, AFDFM, and DSA. In leaf-level severity quantification, LQNet obtained the lowest root mean square error (RMSE) of 1.067 and the highest coefficient of determination (R²) of 0.954 among all methods; the prediction errors were mainly concentrated within ±1% and most densely distributed between 0.2% and 0.7%, and scatter analysis showed the predicted values tightly clustered around the identity line, indicating high stability of the severity estimates. Instance-level analysis further revealed a precision of 84.4%, a recall of 90.7%, and an F1-score of 87.4%, all superior to the compared methods. After mapping the continuous severity estimates onto discrete disease grades, LQNet achieved the highest macro- and micro-averaged accuracies of 97.92% and 97.06%, respectively, indicating reliable class discrimination in both continuous estimation and discrete decision-making; notably, although threshold-based discrete grading narrowed the differences among methods in adjacent grades, LQNet still maintained the best discrimination. Moreover, LQNet required only 38.42M parameters and 30.75G floating-point operations, delivering the best trade-off between performance and computational cost among all models. Overall, the proposed framework effectively converted segmentation outputs into stable, continuous, and repeatable leaf-level severity indicators with high accuracy and efficiency, and demonstrated strong potential for high-throughput phenotyping in foxtail millet blast resistance breeding and genetic association analysis, as well as for practical disease assessment applications. Future work will incorporate larger and more diverse datasets and explore loss functions oriented toward severity quantification to further improve performance and robustness in complex field scenarios.