Ultrasound AI scored 93% on mouse tumours—what would make it useful for patients?
A peer-reviewed model combined stiffness and blood-flow ultrasound maps to classify treatment response in 587 mouse-tumour cases. The result is promising preclinical evidence, not a patient-ready test: there was no human or external validation.
The short version
- What happened
- The peer-reviewed study used 587 paired ultrasound cases from mouse models of breast cancer, sarcoma and melanoma: 175 responsive, 136 stable and 276 non-responsive.
- Who it affects
- If later human studies succeed, an ultrasound-based signal could help oncologists identify ineffective treatment earlier. Today, patients should not interpret the reported 93% as the chance that this model will predict their response: no patient received a decision based on it.
- What remains unproven
- The data came from mouse models under controlled experimental conditions; there were no human participants.
Keep in mind
The data came from mouse models under controlled experimental conditions; there were no human participants.
The study has no locked external hospital or scanner test set, and the accessible methods describe model selection on the full dataset before cross-validation.
The response classes were based on tumour-volume change after experimental treatment, not patient survival, quality of life or treatment benefit.
The 587-case record does not separately disclose the number of individual animals represented in the accessible text.
Source analysis · 2 linked sources. Primary analysed source date: 1 November 2026. Earliest linked supporting source: 23 November 2025.
Read the full report · 4 min The evidence, methods, implications and limitations
Two ultrasound views performed better together in the mouse dataset
Researchers in Cyprus and Germany trained convolutional neural networks on ultrasound images taken before chemo-immunotherapy from mouse models of breast cancer, sarcoma and melanoma. One image type—shear-wave elastography—mapped tissue stiffness. The other—contrast-enhanced ultrasound—used microbubbles to map perfusion. Stiffness and restricted blood flow can both reflect a tumour environment that makes drug delivery and immune response harder, so combining them tests whether two related signals carry more information than either alone. The mice received several chemotherapy, immunotherapy and matrix-targeting regimens, making treatment response the label the model tried to anticipate rather than an image feature measured at the same time.
The peer-reviewed article reports 587 paired cases: 175 tumours later classified as responsive, 136 as stable and 276 as non-responsive using volume-change thresholds adapted from RECIST. A reduction of at least 30% counted as responsive; an increase of at least 20% or new lesions counted as non-responsive, with intermediate changes classified as stable. The combined model reached 93.0% overall test accuracy. Published figures were 87.7% for stiffness alone and 82.5% for perfusion alone; class-specific accuracy for the combined model was 98.3% for responsive, 93.0% for stable and 94.7% for non-responsive cases. Those are classification results within this experimental dataset, not estimates of how accurately a clinic could select treatment for a person.[1][2]
The missing test is a locked model in people and across hospitals
The accessible methods manuscript underlying the article describes images from earlier laboratory experiments, manually selected tumour regions and small colour inputs. Hyperparameters were chosen using the full dataset before ten-fold train, validation and test splits, with eight folds for training and one each for validation and testing. That can estimate internal performance, but it does not provide an untouched external cohort. Images from related experiments may also share scanner, laboratory or treatment characteristics that a model can exploit without learning a broadly transferable biological signal. The accessible record reports 587 tumours or image pairs but does not separately state the number of individual animals represented. It also does not test whether performance holds when tumour outlines are drawn differently, imaging protocols change or a cancer type absent from training is encountered.
For patients, the important question is not whether stiffness and perfusion contain a signal in controlled mouse tumours, but whether a fixed model works prospectively on human cancers imaged at different hospitals. A useful trial would lock the algorithm before enrolment, recruit separate development and external-validation cohorts, report accuracy and calibration by cancer and treatment, and compare predictions with clinicians and established biomarkers. It would also predefine how uncertain predictions are handled and whether poor image quality causes exclusions. The next, harder step would need to show that acting on the result improves outcomes without withholding effective therapy, delaying care or widening access gaps between centres with different ultrasound equipment. Until those tests exist, the study supports a research direction—not a clinical decision tool.[1][2]
The distinction that matters
What each ultrasound input added
The published comparison uses the same 587 paired mouse-tumour cases; percentages are reported test accuracies, not human clinical performance.
Stiffness
Map tissue mechanics with shear-wave elastography
87.7% overall accuracy
Perfusion
Map blood flow with contrast-enhanced ultrasound
82.5% overall accuracy
Combined
Integrate both image types in one model
93.0% overall accuracy