Fully Automated, Semantic Segmentation of Whole-Body ¹⁸F-FDG PET/CT Images Based on Data-Centric Artificial Intelligence

Median DSCs of abdominal organs (10 test datasets) were obtained from models based on 3 separate training subsets: D40 (circle), D20 (square), and D10 (triangle). Inverted triangle (pink) indicates 0.90 mark. Red background highlights organs characterized by low median DSCs (<0.90) and high SDs (Supplemental Fig. 4).

Median DSCs of bone structures (10 test datasets) as obtained from models based on 3 separate training subsets: D40 (circle), D20 (square), and D10 (triangle). Inverted triangle indicates 0.90 mark. Red background highlights bones characterized by low median DSCs (<0.90).

(A) Bar graph demonstrating similar performance of different models for bone segmentation other than forearm bones. Green bar depicts homogeneous training dataset (SMS-20), whereas mixed training datasets (MIX2-18, MIX5-15, and MIX10-10) are represented by red bars. (B) Bar graph showing segmentation performance of forearm bones. Significant performance improvement is seen in mixed training datasets (red bars) compared with homogeneous training dataset (green bar). (C) Forearm bone analysis of individual subject. Images demonstrate that forearm bones are incorrectly segmented in case of SMS20 (green background) model, whereas all mixed models accurately segmented forearm bones (red background).

(A) Error analysis in similarity space for representative lymphoma patient. Horizontal bars depict distance in similarity space, with blue bars characterizing organs with z score of <1.5. Figure shows z scores of >1.5 for liver, kidneys, and bladder (red bars). Corresponding organ segmentations are displayed to right for liver (z = 1.9) and heart (z = 9.9), indicating suboptimal segmentation results that require manual correction. (B) Error analysis in similarity space for representative mesothelioma patient with z scores of >1.5 for liver, heart, bladder, and lung. Incorrect organ segmentations are shown to right for liver (z = 2.8) and heart (z = 8.4). VCI = vena cava inferior.

(A) Organ segmentation of hold-out lymphoma test dataset using training dataset that did not include splenomegaly cases. (B) Organ segmentation of same patient after inclusion of 2 (different) lymphoma datasets and model retraining using expanded training dataset. Updated model was able to recognize new image pattern, resulting in correct segmentation of both liver and spleen.