AAI_2025_Capstone_Chronicles_Combined

12

lesions. The combination of low Dice/Jaccard scores and large, typically positive, volume errors suggests that the current model is not yet suitable for clinical or research use as an automatic volumetry tool. At the same time, the training curves and intermediate validation metrics show that the architecture is capable of learning from this dataset and that further training, more comprehensive cross-validation, and additional regularization strategies would be reasonable next steps. The end-to-end pipeline—from DICOM ingest and RTSTRUCT voxelization through model training, inference, and volumetric calculation—operated reliably, and every case in the dataset could be processed without manual intervention, which was a key methodological objective of this work. 6 ​ Conclusion This project set out to test the hypothesis that accurate slice-wise tumor segmentation on lung CT, combined with DICOM-native geometric reconstruction, can support robust 3D tumor volumetry using clinically drawn RTSTRUCT contours as ground truth. A secondary goal was to develop a reproducible framework that other researchers can use to train and evaluate segmentation-based volumetry models on the Lung1 dataset and similar collections. All experiments were conducted under the seven-week constraints of the capstone module. Given the observed epoch durations and the standard nnU-Net recommendation of training up to 1000 epochs per fold across five folds (Isensee et al., 2021), a fully configured baseline experiment on this dataset, limited to the two-GPU support available in this setup, would likely require substantially longer than the seven-week capstone window. The model evaluated here is therefore best understood as a

​

Figure 5.2​ Volume relative error

A scatter plot of predicted versus ground-truth volume (Figure 5.3) shows that deviations from the ideal y = x line are largest for smaller tumors, where even moderate boundary misalignment translates into large relative error.

​

Figure 5.3​ Predicted vs Ground truth volume

These findings indicate that, for this checkpoint, the model frequently over-segments the tumor and that even modest spatial errors have a large impact on inferred volume, especially in small

409