AI 中文总结
提出MalariAI两阶段解耦框架，通过距离变换分水岭算法实现无标注细胞分割，结合EfficientNet-B0和Focal Loss进行分期分类，在BBBC041数据集上细胞恢复率达75.95%，分类准确率98.36%，并利用Grad-CAM++提供实例级空间证据。
AI 中文摘要
从血涂片显微镜检查中自动诊断疟疾是全球健康AI领域的关键挑战；在资源有限的环境中，缺乏专业显微镜学家仍然是及时准确诊断的主要瓶颈。三个复合故障模式阻碍了现有深度学习系统的可靠临床部署。首先，端到端检测器将未标注细胞视为训练背景，产生的召回率受标注完整性影响较大，而非反映真实的细胞恢复。其次，非极大值抑制倾向于在感染计数最重要的密集涂片区域抑制有效检测。第三，尽管图像级可解释性方法（如Grad-CAM）已应用于疟疾图像分类任务，但现有的全切片检测流程缺乏用于临床审计的每个细胞的空间证据。我们提出MalariAI，一个两阶段解耦框架，在统一流程中解决所有三个故障模式。第一阶段应用标注无关的距离变换引导分水岭算法，在完整的1600x1200血涂片图像中分离每个细胞，在120张图像的NIH BBBC041测试集上通过质心定位恢复了75.95%的真实细胞，无需任何真实标注输入。第二阶段在64x64裁剪块上使用Focal Loss（gamma=2.0，每类逆频率权重）微调EfficientNet-B0，实现了98.36%的整体分类准确率，在罕见的裂殖体和配子体阶段分别达到87.5%和75.0%的每类准确率，而Faster R-CNN基线在同一类别上的AP仅为24.57%和25.95%。每个检测细胞生成的Grad-CAM++热图提供了用于临床审计的实例级空间证据，使显微镜学家能够在个体寄生虫水平验证模型预测，而不会牺牲分类性能。
英文摘要
Automated malaria diagnosis from blood smear microscopy is a critical global health AI challenge; expert scarcity remains the primary diagnostic bottleneck. Existing deep learning systems face three compounding failures: end-to-end detectors treat unannotated cells as background, skewing recall by annotation completeness rather than true cell recovery; Non-Maximum Suppression suppresses valid detections in dense smears; and pipelines lack per-cell spatial evidence for clinical audit. We present MalariAI, a two-stage decoupled framework addressing all three. Stage 1 applies an annotation-agnostic watershed algorithm to isolate every cell in a full 1600x1200 image, recovering 75.95% of ground-truth cells without any ground-truth input. End-to-end, the pipeline reaches a binary parasitized AP@0.5 of 29.10% - the clinically relevant metric for flagging any infected cell - while the stricter multi-class mAP@0.5 of 8.67% mainly reflects watershed's organic region boundaries being penalized against axis-aligned ground-truth boxes, not a localisation failure. Stage 2 fine-tunes EfficientNet-B0 with Focal Loss on ground-truth crops, achieving 98.36% classification accuracy - an oracle upper bound once a cell is correctly localised - with 87.5% and 75.0% accuracy on the rare schizont and gametocyte stages, versus 38.45% and 57.27% AP for a modern YOLOv8s detector evaluated end-to-end on the same classes. Grad-CAM++ heatmaps generated per detected cell provide instance-level spatial evidence for clinical audit; a quantitative energy-in-box analysis confirms this activation is concentrated on the annotated cell body significantly above a geometric chance baseline (+0.0485, paired p = 1.4 x 10^-33), letting microscopists verify predictions at the individual parasite level without sacrificing classification performance.
CommentsSubmitted to Array - Elsivier(under review). 4 authors, includes figures and appendix
