Radiomic prediction of substantial LVSI in endometrial cancer using reduced field of view DWI- a feasibility study
Akiyo Takada1, Daniel A Di Giovanni2, Takuro Horikoshi3
1Department of Diagnostic Radiology, McGill University, Montreal, QC, Canada; Augmented Intelligence & Precision Health Laboratory (AIPHL), Research Institute of the McGill University Health Centre, Montreal, Canada; Diagnostic Radiology and Radiation Oncology, Chiba University Graduate School of Medicine, Chiba, Japan.
Objectives:
The revised FIGO 2023 for endometrial cancer (EC) incorporates lymphovascular space invasion (LVSI) into staging and explicitly designates only substantial/extensive LVSI as adverse. While reduced field-of-view diffusion-weighted imaging (rFOV DWI) improves image quality and staging accuracy for EC, its potential for radiomics-based LVSI assessment remains underexplored. We compared radiomics models based on rFOV DWI and conventional DWI (cDWI) for preoperative prediction of substantial/extensive LVSI and assessed external generalizability.
Materials And Methods:
We included 206 patients: 130 from Hospital A (training/internal test) and 76 from Hospital B (external rFOV DWI test). Whole-tumor radiomics features were extracted from T2WI, rFOV DWI/ADC, and-at Hospital A only-cDWI/ADC. Two calibrated gradient-boosted tree classifiers were developed: rFOV DWI and cDWI models. Discrimination was summarized by AUC with 95 % CIs from stratified bootstrap; generalizability was evaluated in the test sets.
Results:
On the internal test set, the rFOV DWI model achieved an AUC of 0.90 (95 % CI, 0.77-1.00) with sensitivity 1.00, specificity 0.77, and balanced accuracy 0.89; the cDWI model showed an AUC of 0.89 (95 % CI, 0.72-1.00) with sensitivity 0.75, specificity 0.68, and balanced accuracy 0.72. On the external test set, the rFOV DWI model yielded an AUC of 0.74 (95 % CI, 0.56-0.90). Feature importance favored diffusion-related textures; in both models, DWI/ADC features outweighed T2WI.
Conclusion:
The rFOV DWI-based model achieved higher sensitivity without a significant AUC advantage over cDWI. External testing showed a decline in performance despite harmonization. Multicenter validation is required before clinical implementation.
