Cascaded-Recalibrated Multiple Instance Deep Model for Pathologic-Level Lung Cancer Prediction in CT Images

Figure 1

Illustration of patient-level decision-making for the lung cancer diagnosis with multiple nodules. (a) Patient #182 was diagnosed as negative by radiologists but was eventually confirmed as positive. (b) Patient #182 was predicted to be lung cancer positive by the proposed cascaded-recalibrated MIL deep model. (c) Patient #187 was diagnosed as positive by radiologists but was confirmed to be negative by surgical resection. (d) Patient #187 was predicted to be lung cancer negative by the proposed cascaded-recalibrated MIL deep model. “Tex,” “Sph,” and “Mal” are the abbreviations of semantic attributes texture, sphericity, and malignancy, respectively. (a) Patient #182 was diagnosed as lung cancer negative. (b) Patient #182 was predicted to be lung cancer positive. (c) Patient #187 was diagnosed as lung cancer positive. (d) Patient #187 was predicted to be lung cancer negative.