Towards an automated Segment Anything Model (SAM) for lesion segmentation in oncological PET/CT images
Bibliographic record
Abstract
In this work, we test the generalizability of a convolutional neural network, Residual-UNet trained on PET/CT images of one cancer type to other cancer types. We used three oncological PET/CT datasets of different cancer types: lymphoma (n=145), lung cancer (n=168) and melanoma (n=188), collected from two institutions. For each cancer type, the networks were trained to segment the specific single cancer type under 5-fold cross-validation (CV). We evaluated the model on the internal test set of the same cancer type as the training set and then assessed the transferability of the model's lesion segmentation ability on a different cancer type. We further explored different ensembling techniques - Average, Weighted Average, Vote, and STAPLE to combine the five models trained in 5-fold CV as a possible route towards improving model generalizability to new cancer types. For lymphoma-trained ensemble models, we obtained the best Dice similarity coefficient (DSC) (mean, median) of (0.58±0.28, 0.72) on lymphoma test set and a DSC of (0.44±0.25, 0.45) and (0.43±0.31, 0.43) were achieved on lung cancer and melanoma test sets, respectively. Similarly, for lung cancer-trained ensemble models, the best DSC obtained was (0.71±0.20, 0.77) on lung cancer test set and a DSC of (0.41±0.28, 0.48) and (0.42±0.27, 0.48) were achieved on lymphoma and melanoma test sets, respectively. Finally, for melanoma-trained ensemble models, the best DSC obtained was (0.52±0.28, 0.61) on melanoma test set and a DSC of (0.46±0.25, 0.52) and (0.43±0.23, 0.46) were achieved on lymphoma and lung cancer test sets, respectively. We emphasize that ensembling can be a powerful method for generalization, especially in cases when a model is evaluated on a cancer type different from what it was trained on. For internal testing, Weighted Average ensemble generally performed the best, while for testing on a different cancer type, different ensembles performed the best on various training and test set pairs.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.004 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.002 | 0.001 |
| Insufficient payload (model declined to judge) | 0.001 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".