AI for Detecting Abdominal Organ Anomalies in CT Scans
Author Information
Author(s): Sato Junya, Sugimoto Kento, Suzuki Yuki, Wataya Tomohiro, Kita Kosuke, Nishigaki Daiki, Tomiyama Miyuki, Hiraoka Yu, Hori Masatoshi, Takeda Toshihiro, Kido Shoji, Tomiyama Noriyuki
Primary Institution: Osaka University Graduate School of Medicine
Hypothesis
Can free-text radiology reports be used to develop accurate AI models for detecting abnormalities in abdominal CT scans?
Conclusion
The study demonstrates that using free-text radiology reports allows for the development of accurate anomaly detection models without manual annotations.
Supporting Evidence
- The model achieved an AUC of 0.886 for detecting organ abnormalities in the external test cohort.
- Training used a dataset of 66,684 exams from 39,255 patients.
- Models trained on larger datasets with auto-extracted labels outperformed those trained on smaller, human-annotated datasets.
Takeaway
This study shows that AI can help doctors find problems in CT scans of the abdomen by using notes from previous doctors instead of needing special labels.
Methodology
A deep-learning-based pipeline was developed using free-text radiology reports to detect abnormalities in abdominal CT scans across multiple organs.
Potential Biases
Potential selection bias due to exclusion criteria and reliance on radiologist annotations.
Limitations
The study relied on retrospective data and diagnoses made by radiologists rather than final pathological diagnoses, which may misrepresent actual abnormalities.
Participant Demographics
Data collected from 39,255 patients across nine institutions, with self-reported gender.
Statistical Information
Confidence Interval
95% CI: 0.877–0.929
Statistical Significance
p<0.05
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website