Does cross-FOV fusion improve brain metastasis detection in multimodal MRI?

Why two fields of view instead of one

Brain metastases range from lesions smaller than 0.010 mL to masses above 0.500 mL, so a single 3D network faces a severe scale imbalance. A large field of view gives the network surrounding anatomy as context, but a tiny lesion then occupies only a minute fraction of the input volume. A smaller field of view raises that lesion's relative share of the input, at the cost of the context that helps reject look-alike structures.

The paper asks a narrow, controlled question: if the backbone, image resolution and input modalities are held fixed, does combining predictions from two different spatial fields of view improve lesion-level detection? The authors frame this as separating the effect of spatial context from the effects of resolution and modality choice, which are usually changed together in multiscale pipelines.

Weighted late fusion of whole-volume probability maps

The framework trains two compact 3D U-Nets independently, one on 96³ inputs and one on 64³ inputs. Each produces a whole-volume probability map. Fusion happens after prediction, at the probability level: the two maps are combined with a weight α that sets the share of the 96³ map, and the fused map is then thresholded and filtered.

Because the networks are trained separately and merged only at the output, the fusion step adds no joint training and is computationally light. It also makes the field of view the only variable deliberately varied between the two branches, which is what allows the comparison below to be read as an effect of spatial context.

A smaller field of view alone did not help

Reducing the field of view in isolation did not beat the 96³ baseline. The 64³ model's best observed F1 was 0.413 at probability threshold 0.70, with sensitivity 0.484, precision 0.360 and 5.907 false positives per patient. An independently trained 64³ replicate reached F1 0.433 at the same threshold, with sensitivity 0.541, precision 0.361 and 6.588 false positives per patient.

A sampling ablation pointed the same way. At roughly matched false-positive burden, baseline sensitivity was 0.544 at 4.00 false positives per patient, versus 0.501 at 3.99 with weighted sampling, and there was no consistent size-stratified advantage. Sampling more often, or shrinking the field of view, was not sufficient on its own.

Fusion improved precision and F1 at intermediate weights

Intermediate fusion weights gave the best precision–sensitivity balance. At α = 0.60, fusion reached sensitivity 0.582, precision 0.602, F1 0.592 and 2.639 false positives per patient. Relative to the 96³ endpoint, that is a 17.2% relative F1 increase and a 57.2% reduction in false positives per patient.

A patient-level bootstrap with 10,000 resamples put fused F1 at 0.592 (95% CI 0.538–0.643) against 0.505 (95% CI 0.428–0.583) for the 96³ model. The paired F1 difference was +0.087 (95% CI 0.024–0.150; p = 0.003), with 3.526 fewer false positives per patient (95% CI 1.691–5.722). The authors chose α = 0.60 over an equally rounded-F1 setting at α = 0.70 because it detected 26 additional true lesions.

The advantage was not tied to one operating point. When each method selected its own best-F1 probability threshold, fusion reached F1 0.593 at threshold 0.65 with sensitivity 0.559 and 2.26 false positives per patient, while the 96³ model reached F1 0.515 at threshold 0.55 with sensitivity 0.624 and 5.49 false positives per patient.

The gain is not just ensemble averaging

Averaging two independently trained networks can improve results by itself, so the paper adds a same-field-of-view control. A second 64³ model (seed 43) was trained and its probability map averaged 50/50 with the original 64³ model. At each method's best-F1 threshold, this 64³ + 64³ ensemble achieved sensitivity 0.523, precision 0.472, F1 0.496 and 4.021 false positives per patient.

The cross-field-of-view 96³ + 64³ ensemble reached sensitivity 0.559, precision 0.630, F1 0.593 and 2.258 false positives per patient. In a paired patient-level bootstrap, cross-field-of-view fusion improved F1 by 0.096 (95% CI 0.059–0.132; p < 0.001), improved precision by 0.158 (95% CI 0.113–0.203; p < 0.001) and reduced false positives by 1.763 per patient (95% CI 1.247–2.340; p < 0.001). The sensitivity difference of +0.036 was not significant (95% CI −0.008–0.075; p = 0.114).

The authors read this as evidence that generic ensembling contributes to performance, while the cross-field-of-view result adds a precision, F1 and false-positive benefit that averaging alone does not explain. That interpretation rests on the control comparison, not on a direct measurement of what each branch contributes.

Small metastases remain the hard case

Stratifying the selected α = 0.60 fusion model by ground-truth lesion volume shows sensitivity rising monotonically with size: 0.057 below 0.010 mL (4 of 70 lesions), 0.326 at 0.010–0.025 mL (31 of 95), 0.414 at 0.025–0.050 mL (36 of 87), 0.583 at 0.050–0.100 mL (49 of 84), 0.698 at 0.100–0.500 mL (97 of 139) and 0.891 at 0.500 mL or more (171 of 192).

Crucially, under matched post-processing, fusion did not improve sensitivity within individual size strata relative to the 96³ baseline. Its benefit was improved precision and F1 through false-positive suppression. Readers comparing this work with multiscale methods that claim better small-lesion sensitivity should note that this framework's gain sits on the false-positive side of the trade-off, not the detection side.

The 0.010 mL filter is an empirical trade-off

Holding the fusion weight at α = 0.60 and the probability threshold at 0.60, the authors swept the minimum predicted-component volume over 0, 0.0025, 0.005, 0.010, 0.025 and 0.050 mL. The 0.010 mL filter gave the highest overall F1, 0.592. Removing component filtering raised overall sensitivity from 0.582 to 0.619 and sensitivity for lesions below 0.010 mL from 0.057 to 0.171, but pushed false positives from 2.64 to 6.32 per patient and dropped F1 to 0.488.

Tightening the filter to 0.025 or 0.050 mL reduced false positives further, to 1.69 and 1.10 per patient, but also lowered F1 to 0.572 and 0.533. The chosen setting is therefore a measured compromise between very-small-lesion sensitivity and false-positive suppression, not a free improvement.

Cross-FOV agreement filtering trades sensitivity

As an exploratory post-hoc control, the authors kept the 60/40 fusion at probability threshold 0.60 and required each connected-component candidate to receive a minimum mean support from both field-of-view branches. At an agreement threshold of 0.35, false positives fell from 2.639 to 2.268 per patient and precision rose from 0.602 to 0.630, while sensitivity fell from 0.582 to 0.561. F1 was essentially unchanged (0.592 versus 0.593; paired-bootstrap difference +0.001, 95% CI −0.009–0.012; p = 0.786).

The lesson is that requiring two fields of view to agree reduces false positives mainly by discarding borderline true lesions. It does not add a second, independent gain on top of weighted fusion.

What the residual errors and external test show

An appearance-based audit of the 40 highest-confidence unmatched predictions found 23 (57.5%) compact focal, lesion-like candidates; 8 (20.0%) peripheral, boundary-associated or anatomical structures; 7 (17.5%) large diffuse or irregular non-lesional regions; and 2 (5.0%) near-ground-truth cases lost to split/merge or one-to-one matching. The audit is descriptive, not radiologist-adjudicated, and the proportions apply only to that high-confidence subset. The authors conclude that false-positive suppression needs more than threshold adjustment.

Scope matters here. The main results come from a 97-patient development cohort. A descriptive analysis of 509 non-DEV97 cases drew from the same BraTS-METS training dataset and used patch-level rather than patient-level splitting, so it is not an independent held-out test. The official BraTS evaluation showed a substantial internal-to-external gap, with an ET instance F1 of 0.385. The fusion benefit documented here is therefore a development-cohort finding about the precision–false-positive trade-off, not a demonstrated generalization gain.

Where cross-FOV fusion earns its place

The paper's practical priority is clear: when the backbone, resolution and modalities are fixed and false positives are the binding problem, weighted late fusion of 96³ and 64³ probability maps is a cheap way to move the operating point. At α = 0.60 it delivered F1 0.592 with 2.639 false positives per patient, and the same-field-of-view control shows the gain exceeds what ensembling alone provides.

The boundary is equally specific. Fusion did not raise sensitivity within any lesion-size stratum, and the smallest lesions stayed near 0.057 sensitivity below 0.010 mL. A team whose clinical concern is missing tiny metastases should treat cross-FOV fusion as a false-positive control that must be paired with something else, and should weigh the 0.010 mL component filter as a tunable trade-off rather than a default.

Sources used in this answer