Hierarchical Classification of Pulmonary Lesions: A Large-Scale - - PowerPoint PPT Presentation

Hierarchical Classification of Pulmonary Lesions: A Large-Scale Radio-Pathomics Study

Oct 2020 Jiancheng Yang*, Mingze Gao*, Kaiming Kuang, Bingbing Ni, Yunlang She, Dong Xie, Chang Chen.

Introduction – Methodology – Experiments – Conclusion

Introduction How to reduce the annotation ambiguity from human experts?

Probabilistic Radiomics: Ambiguous Diagnosis with Controllable Shape Analysis. Jiancheng Yang et al. MICCAI'19

Medical image annotation is very ambiguous due to

• Expertise of annotators
• Imaging issues
• Inherent issues
• …

On LIDC-IDRI, the largest public database for lung nodule diagnosis,

• nly ~1,183 out of 2,635 (<50%)

nodules are diagnosed consistently by 4 radiologists.

Introduction – Methodology – Experiments – Conclusion

Introduction Use pathologically confirmed labels (clinical golden standard)!

Introduction – Methodology – Experiments – Conclusion

Introduction Use pathologically confirmed labels (clinical golden standard)! Our in-house dataset Pulmonary-RadPath

• Large-scale
• Hierarchical
• Multi-Disease

Introduction – Methodology – Experiments – Conclusion

Methodology: Pulmonary-RadPath

All data is collected from a single clinical center:

Shanghai Pulmonary Hospital, Tongji University, Shanghai, China.

Introduction – Methodology – Experiments – Conclusion

Methodology: Hierarchical Classification

Take-Home:

• Learn each hierarchy with

separate head

• Keep a Leaky Node as a

virtual class for “others”

• Make densely connected

hierarchies

Introduction – Methodology – Experiments – Conclusion

Experiments

• H1: Cancer / Non-Cancer
• H2: Cancer subtypes
• H3: Non-Cancer subtypes
• H4: Adenocarcinoma

subtypes: invasive (IA) / non-invasive (non-IA) H4 prior arts: 88.0 / 92.0 / 92.1

3D Deep Learning from CT Scans Predicts Tumor Invasiveness of Subcentimeter Pulmonary Adenocarcinomas Wei Zhao*, Jiancheng Yang*, Yingli Sun, Cheng Li, Weilan Wu, Liang Jin, Zhiming Yang, Bingbing Ni, Pan Gao, Peijun Wang, Yanqing Hua, Ming Li. Cancer Research (IF-9.130), 2018

Introduction – Methodology – Experiments – Conclusion

Conclusion

• We build a large-scale in-house Radio-Pathomics dataset, named

Pulmonary-RadPath: 5,134 radiological CT images with pathologically confirmed labels.

• We develop the first hierarchical multi-disease classification system
• f pulmonary lesions.
• The discriminative performance is impressive compared to prior arts.
• Limitation:
• Data bias: pathological analysis is performed on high-risk subjects.
• Long-tailed: certain disease types are naturally very long-tailed.

Thanks for Listening

Check out my home page for materials on this study https://jiancheng-yang.com/ WeChat jekyll4168