First-hand information for everyone
Main discovery The authors show that tumour appearances learned from pre‑operative MR images can guide localisation on intra‑operative CT without paired MR–CT training images. This enables cross‑modality tumour localisation in a zero‑shot setting using weak supervision (no voxel-wise CT tumour maps required). (F1)
How the method works A tumour‑presence classifier was trained on pre‑operative MR volumes using only binary labels (tumour present/absent), not voxel‑wise delineations. (F2)
At inference, the MR classifier is applied to fixed‑size local windows extracted from the CT image. The authors hypothesise — and use — the fact that tumour‑containing windows tend to receive higher tumour‑presence probabilities than tumour‑free windows. These window probabilities form the weak supervision signal. "Weak supervision" here means using a lower‑resolution or coarser label (window‑level presence) instead of precise voxel‑level tumour outlines. (F3)
A reinforcement‑learning (RL) agent moves the window through the CT volume. Actions translate the window by a fixed step along image axes. The scalar reward equals the change in the classifier’s tumour‑presence probability between successive windows. The classifier weights are kept fixed while the RL policy is optimised. (F4)
Each visited window contributes its probability to the voxels it covers. For each voxel, the maximum probability from all windows that covered it is retained to form a voxel‑wise tumour score map. That map is thresholded (threshold chosen by grid search) to produce a segmentation. (F5)
Three transfer strategies were evaluated. Zero‑shot transfer applies the MR classifier directly to CT windows. Few‑shot adaptation fine‑tunes the MR classifier on a small CT set. Synthetic transfer first translates CT into synthetic MR using an unpaired CycleGAN trained on separate MR and CT development sets, then applies the MR classifier to the synthetic MR. (F6)
Data and baselines MR training data came from 334 patients. CT development data came from 313 patients. Evaluation used an external HCC cohort (TCIA HCC‑TACE‑Seg, 105 patients) with radiologist voxel‑wise tumour annotations. (F7)
Baselines and ablations included two CT‑only RL variants: one with full supervision using rewards from tumour labels, and one with weak supervision using a CT‑only classifier. We also evaluated a fully supervised (FS) nn‑UNet baseline. Registration methods were included for contextual comparison; their reported results used different evaluation cohorts and are not directly comparable. Few‑shot adaptation was fine‑tuned on 40 CT development volumes. (F8)
Concrete results On the external HCC test set, synthetic transfer achieved the highest mean patient‑level Dice of 0.31 ± 0.10 and a localisation success rate of 93%. The authors report that synthetic transfer’s Dice was significantly higher than all other tested methods (all p < 0.01). (F9)
Other reported results (mean Dice ± SD; success rate): - Zero‑shot transfer: 0.16 ± 0.06; 67% success. (F10) - Few‑shot transfer: 0.19 ± 0.09; 72% success. (F10) - CT‑only nn‑UNet (fully supervised): 0.24 ± 0.12; 81% success. (F10)
Significance and caveats The results suggest that reducing the MR–CT modality gap (via synthetic MR) strengthens guidance from an MR‑trained classifier and can improve localisation on intra‑operative CT compared with CT‑only approaches. (F11)
However, despite a 93% localisation success rate for synthetic transfer, the mean Dice of 0.31 indicates that accurate HCC boundary delineation on intra‑operative CT remains challenging. The authors note this limitation and suggest alternative forms of supervision could be explored to improve boundary accuracy. (F11)
Reported comparisons to deformable registration methods are provided as contextual references only. Those registration results were obtained on different evaluation cohorts in the cited works, which limits direct comparability. (F12)