Report-Supervision Loss (R-Super) for CT Tumor Segmentation

• The paper introduces R-Super, a weakly supervised training paradigm that leverages radiology reports to guide CT tumor segmentation when voxel-level tumor masks are scarce.
• It employs a dual loss strategy—Volume Loss for enforcing organ-level tumor volume consistency and Ball Loss for dynamic pseudo-mask generation based on lesion count, size, and location.
• The method shows significant performance gains over mask-only training, especially in low- and medium-mask regimes, by directly integrating report-derived spatial and anatomical information.

Searching arXiv for the specified paper to ground the article in the current record. Report-Supervision Loss (R-Super) is a weakly supervised training paradigm for CT tumor segmentation that uses radiology reports as direct supervision when voxel-level tumor masks are scarce. Introduced in "Learning Segmentation from Radiology Reports" (Bassi et al., 8 Jul 2025), it addresses a practical asymmetry in clinical data availability: manual 3D tumor masks are expensive and rare, whereas hospitals routinely store large numbers of CT–report pairs. R-Super converts free-text reports into spatial constraints on the segmentation output itself, rather than using reports only for auxiliary classification or CLIP-like contrastive learning. Its core mechanism is a pair of report-derived losses—a deep-supervision Volume Loss and a final-output Ball Loss—that translate report content such as tumor count, location, and size into anatomically constrained optimization signals for segmentation.

1. Problem setting and conceptual scope

R-Super is formulated for weakly supervised segmentation with two kinds of training data: a relatively small set of CT–Mask pairs with voxel-wise tumor annotations, and a much larger set of CT–Report pairs with no tumor masks. The motivating observation is that public abdominal CT segmentation datasets typically have from tens to low thousands of masks per tumor type, whereas a single hospital may have thousands to hundreds of thousands of reports. Within this setting, reports are treated as a scalable source of supervision.

The central distinction between R-Super and earlier report-based strategies is methodological. Prior report-based methods mostly used reports indirectly, for example by extracting image-level labels for multitask learning or by doing CLIP-like contrastive learning. The authors argue that this indirect supervision does not directly optimize segmentation quality. R-Super instead uses report content to penalize false positives and false negatives in segmentation itself.

The report attributes used for supervision are clinically specific. Reports often contain whether a tumor exists, how many tumors there are, which organ or sub-region they occupy, and their diameters. R-Super uses these attributes as direct constraints on segmentation. This suggests a shift from semantic supervision at the image level to geometry-aware supervision at the voxel level, while still operating without manual tumor masks for the report-only cases.

2. Report parsing and anatomical priors

The conversion from free text to supervision is mediated by a zero-shot LLM. The paper uses Llama 3.1 70B AWQ, prompted with radiologist-designed prompts, to extract three structured fields from reports: tumor count, tumor location, and tumor size (Bassi et al., 8 Jul 2025). Location can be at organ level or organ sub-segment level; the paper highlights pancreas sub-segments (head, body, tail) and kidneys.

Tumor size is converted into approximate volume by explicit heuristics. If only one diameter $d_1$ is reported, the tumor is treated as a sphere with volume

$\frac{d_1^3 \pi}{6}.$

If three diameters $d_1,d_2,d_3$ are reported, an ellipsoid approximation is used:

$\frac{d_1 d_2 d_3 \pi}{6}.$

If two diameters $d_1,d_2$ are reported, the third is imputed as

$d_3 = \frac{d_1+d_2}{2},$

and the ellipsoid formula is then used. For each organ or organ sub-segment $o$, the volumes of all tumors mentioned in that region are summed to obtain the report-derived total tumor volume $V_{r,o}$.

R-Super also requires organ masks as anatomical priors. These are precomputed binary segmentations of organs or sub-segments, used to define spatial support where the report indicates tumor presence. The authors used an nnU-Net trained on AbdomenAtlas 2.0 to segment 39 organs and anatomical structures, including pancreas sub-segments and kidneys. These masks are not the target tumor masks; they specify where report-derived supervision should be applied. The paper notes that sub-segment masks improve performance but are not strictly necessary.

3. Volume Loss

The Volume Loss provides deep supervision at an intermediate decoder stage. An intermediate decoder feature map is converted into a one-channel tumor probability map by a $1\times1\times1$ convolution followed by sigmoid or softmax and interpolation to CT spacing. This intermediate map is supervised by an organ-wise volume consistency term plus a background suppression term.

Let $\mathbf{T} = [t_{h,w,l}]$ denote the tumor probability map, $\frac{d_1^3 \pi}{6}.$0 the binary organ mask, and $\frac{d_1^3 \pi}{6}.$1 the voxel volume determined by CT spacing. The segmented tumor volume inside organ $\frac{d_1^3 \pi}{6}.$2 is computed as a soft volume:

$\frac{d_1^3 \pi}{6}.$3

Because this sums probabilities rather than thresholded predictions, it functions as a differentiable estimate of total tumor burden within the organ.

To compare $\frac{d_1^3 \pi}{6}.$4 with the report-derived volume $\frac{d_1^3 \pi}{6}.$5, the paper defines a normalized discrepancy:

$\frac{d_1^3 \pi}{6}.$6

with $\frac{d_1^3 \pi}{6}.$7 as a stability constant. The paper assigns two roles to this constant: improving numerical stability and ensuring that the loss minimum occurs at $\frac{d_1^3 \pi}{6}.$8 when $\frac{d_1^3 \pi}{6}.$9.

Because report-derived volumes are noisy, the loss incorporates a tolerance band. With tolerance parameter $d_1,d_2,d_3$0, set to $d_1,d_2,d_3$1, the foreground term is

$d_1,d_2,d_3$2

This makes the loss zero when predicted volume is sufficiently close to the report-estimated volume.

The background term suppresses tumor probability outside the reported organ or sub-segment:

$d_1,d_2,d_3$3

The final Volume Loss is

$d_1,d_2,d_3$4

Operationally, this loss enforces two conditions: the right total amount of tumor should appear inside the reported organ, and tumor probability should remain low elsewhere. Because it is applied at an intermediate layer, it acts as a permissive global structural constraint rather than an exact boundary supervisor.

4. Ball Loss and dynamic pseudo-mask generation

The Ball Loss is applied to the final segmentation output and is explicitly stricter than the Volume Loss. Its aim is to make the prediction agree with the report not only in total volume and location, but also in tumor count and approximate lesion diameters. Unlike the Volume Loss, it is described algorithmically rather than as a single closed-form equation.

For each organ $d_1,d_2,d_3$5, the final tumor probability map is first masked by the organ segmentation, yielding $d_1,d_2,d_3$6. For each tumor described in the report, starting from the largest, the algorithm searches for the most likely lesion location using a Ball Convolution. The kernel is fixed and non-learnable, with diameter equal to the reported lesion diameter plus a small margin. The paper first explains the intuition with a binary spherical kernel, but the actual implementation uses a center-weighted ball kernel: the value is $d_1,d_2,d_3$7 at the center and decays according to a 3D Gaussian with standard deviation $d_1,d_2,d_3$8, where $d_1,d_2,d_3$9 is the ball diameter, and is zero outside the ball. Convolution uses stride $\frac{d_1 d_2 d_3 \pi}{6}.$0, zero padding, and odd-sized kernels so that input-output alignment is preserved.

Sliding this kernel over the masked output produces, at each candidate center, a score corresponding to the summed tumor probabilities in a ball-shaped neighborhood, weighted toward the center. The position of maximal response is taken as the highest-probability ball for a tumor of that size. Once that ball is localized, the method selects the top-$\frac{d_1 d_2 d_3 \pi}{6}.$1 most probable voxels inside it, where $\frac{d_1 d_2 d_3 \pi}{6}.$2 is the tumor’s estimated volume converted to voxels. These top-$\frac{d_1 d_2 d_3 \pi}{6}.$3 voxels are written into a blank mask as a pseudo-label for that tumor, and those voxels are simultaneously zeroed out in the model output so that subsequent tumors cannot reuse them. The procedure then repeats from the next-largest reported tumor to the smallest.

This greedy assignment ensures that the dynamically generated pseudo-mask matches the report in lesion count, organ or sub-segment location, approximate size, and total volume (Bassi et al., 8 Jul 2025). Because only the top-$\frac{d_1 d_2 d_3 \pi}{6}.$4 most probable voxels inside a plausible ball are used, the resulting pseudo-label can have any shape within that neighborhood; it is not forced to be spherical. The paper emphasizes that this design avoids over-penalizing uncertain tumor borders and focuses learning more strongly on lesion centers.

The Ball Loss then uses these dynamic pseudo-masks as labels under standard segmentation losses—cross-entropy and Dice. The text states that, to accommodate uncertainty and imprecise diameters in reports, the cross-entropy term is weighted more heavily on tumor voxels with higher predicted tumor probability, and a small border margin is not penalized. The paper does not provide an explicit closed-form equation for this weighted boundary-tolerant Ball Loss, but the operational pipeline is specified: parse report, estimate lesion diameters and volumes, localize candidate lesion balls by convolution, extract top-$\frac{d_1 d_2 d_3 \pi}{6}.$5 voxels per lesion, assemble a pseudo-mask, and optimize the final prediction against that pseudo-mask.

5. Training objective, protocol, and datasets

The full objective combines supervised segmentation losses on true masks with report-derived losses on report-only cases. For CT–Mask pairs, the network is trained with traditional segmentation losses, specifically cross-entropy and DSC or Dice. For CT–Report pairs, the Volume Loss is applied at an intermediate layer and the Ball-derived pseudo-mask supervision is applied at the final output using standard segmentation losses. The paper gives the relative weighting at training time: segmentation losses receive weight $\frac{d_1 d_2 d_3 \pi}{6}.$6 and “our losses” receive weight $\frac{d_1 d_2 d_3 \pi}{6}.$7.

The paper does not present a single master formula, but it gives a wording from which a faithful reconstruction follows. For a mask-labeled sample,

$\frac{d_1 d_2 d_3 \pi}{6}.$8

For a report-only sample,

$\frac{d_1 d_2 d_3 \pi}{6}.$9

where $d_1,d_2$0 is the dynamic pseudo-mask generated by the Ball Loss procedure. A dataset-level objective consistent with the paper’s wording is

$d_1,d_2$1

Training proceeds in two stages. First, the model is trained only on CT–Mask pairs using standard segmentation losses. Second, it is fine-tuned jointly on both CT–Mask pairs and CT–Report pairs. This staging is important because the report-derived constraints are weak and approximate; pretraining on true masks provides a strong initialization so that report-based losses refine rather than bootstrap the model from scratch. The experiments use MedFormer, a transformer-CNN hybrid segmentation network, although the method is intended to be architecture-agnostic.

Several practical heuristics are integral to training. Since learning is patch-based, each training patch fully covers one target organ. If a second organ with a reported tumor is only partially present in the patch, no report-based loss is applied for that second organ. If a report mentions a tumor but does not mention tumor size, the losses are not applied for that organ; these cases were under $d_1,d_2$2 of reports. The paper also notes that when the model output is nearly zero everywhere, the Ball Loss alone may struggle, because it can only amplify the highest-probability region even if all probabilities are tiny; the Volume Loss supplies stronger gradients in this regime. The number of masks and reports was balanced with data augmentation, and a random $d_1,d_2$3 hold-out validation split was used.

Dataset	Composition	Role
UCSF-Train	6,718 CT–Report pairs without masks: 2,229 pancreatic tumor cases, 2,738 kidney tumor cases, 1,751 normal cases	report-only supervision
AbdomenAtlas 2.0	9,262 CT–Mask pairs from 88 hospitals across 19 countries, including 344 pancreatic tumor CTs and 1,674 kidney tumor CTs	public masked dataset
AbdomenAtlas-Small	50 pancreatic tumor CTs, 50 kidney tumor CTs, 100 normal CTs	low-label subset
UCSF-Test	169 kidney tumor CTs, 139 pancreatic tumor CTs, 100 normal CTs	internal evaluation
JHH-Test	50 pancreatic tumor CTs, 50 normal CTs	external evaluation

In the main report-based experiments, pancreas models used 2.2K pancreatic reports and kidney models used 2.7K kidney reports.

6. Empirical behavior and ablation findings

The principal empirical claim is that report supervision improves performance over mask-only training in low-, medium-, and higher-mask regimes (Bassi et al., 8 Jul 2025). The paper repeatedly highlights F1-score gains: compared with standard segmentation trained without reports, R-Super improved tumor detection F1-score by up to about $d_1,d_2$4 in the medium-mask regime, $d_1,d_2$5 in the few-mask regime, and $d_1,d_2$6 even when 1.7K kidney masks were available.

For pancreas with 344 training masks plus 2.2K reports, R-Super outperformed all baselines. On JHH-Test, pancreas Dice improved from 51 to 59, NSD from 59 to 69, sensitivity from 82 to 91, specificity from 84 to 92, F1 from 78 to 94, and AUC remained high at 88 versus 88 for the MedFormer segmentation baseline. On UCSF-Test, pancreas F1 improved from 62 to 83 and AUC from 85 to 89. In the low-mask pancreas setting with only 50 masks, improvements remained substantial: on JHH-Test, Dice increased from 38 to 49, NSD from 41 to 52, sensitivity from 62 to 67, specificity from 63 to 75, F1 from 62 to 68, and AUC from 62 to 64; on UCSF-Test, pancreas F1 increased from 47 to 73.

For kidney, the gains were more modest but still present, especially in the many-mask regime. With 1.7K kidney masks plus 2.7K reports on UCSF-Test, R-Super improved Dice from 71 to 75, NSD from 73 to 78, sensitivity from 65 to 80, and specificity from 68 to 70. In the few-mask kidney setting with 50 masks, performance was mixed depending on metric: Dice 64 to 65, NSD 68 to 66, sensitivity 68 to 73, specificity 63 to 55. The paper summarizes this as matching segmentation with few masks and surpassing it with many masks for kidney.

The external validation result is central to the paper’s argument about generalization. R-Super was trained using UCSF reports but tested on JHH, an unseen hospital with different acquisition characteristics. The paper emphasizes that report supervision still improved pancreas performance there, with up to $d_1,d_2$7 F1-score in external validation. This suggests that report-derived supervision is not merely overfitting to the source institution’s reporting style.

Results stratified by tumor size indicate benefits for both small and large tumors. For small pancreatic tumors ($d_1,d_2$8 cm) with 344 masks, on JHH-Test Dice rose from 17 to 25, NSD from 36 to 48, sensitivity from 65 to 80, specificity from 83 to 89, F1 from 67 to 93, and AUC remained 88. On UCSF-Test, F1 improved from 58 to 77 and AUC from 85 to 89. For large pancreatic tumors ($d_1,d_2$9 cm), on JHH-Test Dice rose from 18 to 33, NSD from 33 to 51, and F1 from 82 to 94. The authors interpret these results as evidence that count, size, and location constraints assist not only with obvious masses but also with earlier, smaller lesions.

Ablation results indicate that both report-derived components matter and that their combination is strongest. On pancreas with 344 masks and 2.2K reports, using only Volume Loss gave JHH Dice/NSD of 59/61 and UCSF-Test F1/AUC of 80/82; using only Ball Loss gave 59/59 and 64/90; full R-Super achieved 59/69 and 83/89. This suggests that Volume Loss is particularly useful for recall and broad volumetric guidance, while Ball Loss sharpens localization and final instance structure. The paper also reports that pancreas sub-segment masks materially improve performance: without them, using only whole-pancreas masks, F1/AUC drop to 88/84 on JHH and 74/80 on UCSF.

7. Assumptions, limitations, and interpretation

R-Super is built around several explicit assumptions. First, it depends on reports mentioning useful tumor attributes, especially size. If a report mentions a tumor but not tumor size, the losses are not applied for that organ, although these cases were fewer than $d_3 = \frac{d_1+d_2}{2},$0 of reports. Second, report-derived volumes are approximate because diameters may be imprecise and sphere or ellipsoid approximations are crude. The tolerance parameter in the Volume Loss is a direct response to this uncertainty.

Third, the method assumes availability of organ masks or an organ segmentation model, and performance benefits from sub-organ segmentation when reports localize tumors to sub-segments. Fourth, the Ball Loss uses a greedy assignment procedure that may be imperfect when multiple lesions are close together or when predicted probabilities are very poor. Fifth, because the report parser is based on an LLM, extraction errors can propagate into supervision. The paper cites $d_3 = \frac{d_1+d_2}{2},$1 accuracy for extracting tumor presence and location from reports, but not a full lesion-by-lesion extraction accuracy for all attributes. Sixth, the pseudo-mask generation depends on current model predictions, so early training stability relies on pretraining with true masks and on the supplementary role of the Volume Loss (Bassi et al., 8 Jul 2025).

A common misconception would be to treat R-Super as merely an image–text pretraining strategy or a form of report-based classification regularization. The method is more specific: it translates free-text radiology reports into direct segmentation supervision by combining global organ-level volume consistency with local lesion-level pseudo-label generation. Another misconception would be that the Ball Loss enforces spherical tumors. The paper explicitly states the opposite: the pseudo-label can take any shape within the selected neighborhood, because only the top-$d_3 = \frac{d_1+d_2}{2},$2 most probable voxels inside the ball are retained.

In aggregate, R-Super defines a framework in which free-text reports become a supervisory signal for voxel-wise learning. The Volume Loss provides a soft, anatomically constrained estimate of how much tumor should exist in each organ, while the Ball Loss transforms lesion count, size, and location into dynamic lesion-wise pseudo-masks aligned to the model’s current predictions. The reported results indicate that this combination allows segmentation models to learn from thousands of report-only studies and to improve over mask-only training, especially when manual masks are scarce but also when hundreds to thousands of masks are available.