ResNet50-U-Net Deep Learning Model Segments Brain Tumor Subregions from Multimodal MRI

Researchers have unveiled a deep learning framework that pairs a ResNet50 residual encoder with a U-Net-style decoder to segment brain tumor subregions from multimodal magnetic resonance imaging, and the results reveal a nuanced picture of what residual feature extraction can and cannot deliver in this demanding clinical task. The study, published in Discover Artificial Intelligence, reports that the integrated ResNet50-U-Net model achieved the highest Dice Similarity Coefficient scores for Whole Tumor and Enhancing Tumor segmentation among the architectures tested, reaching 0.9100 and 0.8151 respectively on the BraTS2020 benchmark. Yet the story is not one of uniform dominance: a far lighter convolutional neural network baseline outperformed the deeper model on Tumor Core segmentation, recording a Dice score of 0.7915, a finding that underscores how architecture, tumor subregion, and computational budget interact in ways that resist simple rankings.

The clinical motivation behind the work is substantial. Brain tumors remain among the most critical health conditions worldwide, and gliomas in particular demand precise delineation for diagnosis, treatment planning, and disease monitoring. Magnetic resonance imaging is the standard tool for detecting these tumors, but manual segmentation is complicated by the soft tissue environment, the irregular and variable shapes tumors adopt, and the ambiguity of their location and size. Tumors themselves span a spectrum from benign, slow-growing masses with distinct boundaries that can be surgically removed, to malignant growths that spread quickly into adjacent brain regions with irregular borders, and secondary metastatic tumors that originate elsewhere in the body. Each MRI sequence highlights different aspects of the pathology, which is why the researchers combined four modalities rather than relying on any single scan.

The technical centerpiece of the framework is a learnable one-by-one convolutional modality adapter. Standard ResNet50 networks are built for three-channel natural image input, but the segmentation task requires fusing FLAIR, T1, T1ce, and T2 sequences, each capturing complementary anatomical and pathological characteristics of gliomas. The adapter transforms the four-channel input into a three-channel representation compatible with the ResNet50 encoder, allowing the researchers to initialize the encoder with ImageNet pre-trained weights while preserving multimodal MRI information for dense pixel-wise prediction. The classification head of the encoder was removed, since the goal is segmentation rather than image-level classification, and intermediate feature maps from different encoder stages were passed to the decoder through skip connections so that deep semantic information could be combined with lower-level spatial details critical for preserving tumor boundaries and small enhancing regions.

The experimental design emphasized fairness and control. The team worked with 368 complete labelled cases from the BraTS2020 training dataset, split patient-wise into 257 training, 55 validation, and 56 independent test cases to prevent data leakage between sets drawn from the same patient. After preprocessing and tumor slice selection, this yielded 17,053 training slices, 3,573 validation slices, and 3,706 test slices. Each input sample was a 224 by 224 by 4 multimodal MRI slice, and each output a 224 by 224 by 3 segmentation map corresponding to the three clinically relevant subregions: Whole Tumor, representing the complete tumor-affected area; Tumor Core, the central mass; and Enhancing Tumor, the contrast-enhancing active component. Intensity normalization reduced inter-scan variation, and only lightweight random flipping was used for augmentation, with aggressive transformations like large rotations and heavy zooming deliberately avoided to protect anatomical consistency.

All three models, the lightweight CNN, the standard U-Net, and the proposed ResNet50-U-Net, were trained under identical conditions: the same patient-wise split, preprocessing pipeline, input resolution, output format, loss function, threshold-tuning method, and evaluation metrics. The loss function itself was a hybrid, combining weighted binary cross-entropy with weight 0.45, Dice loss with weight 0.35, and focal Tversky loss with weight 0.20. Weighted binary cross-entropy improves pixel-wise classification of tumor versus non-tumor classes, Dice loss improves overlap between predicted and ground-truth masks, and focal Tversky loss mitigates the class imbalance that plagues smaller tumor regions like the Enhancing Tumor. Training ran for up to 60 epochs with early stopping and learning-rate reduction based on validation performance, using TensorFlow/Keras with GPU acceleration and mixed precision.

The headline results show the ResNet50-U-Net achieving the best Dice and IoU scores for Whole Tumor segmentation, along with the lowest HD95 boundary error of 8.52 for that region. For Enhancing Tumor, it again led with a Dice score of 0.8151, an IoU of 0.6880, and the lowest HD95 value of 2.53. The U-Net baseline delivered solid intermediate performance across all three subregions, with Dice scores of 0.8982, 0.7872, and 0.8124 for WT, TC, and ET respectively. But the compact CNN, with only 108,211 parameters and a 0.41 MB weight file, proved remarkably competitive on Tumor Core, posting the best Dice of 0.7915, the best IoU of 0.6549, and the lowest HD95 of 5.04 for that region. The proposed model, by contrast, contained 32,603,682 parameters and weighed in at 124.37 MB, a substantial computational footprint.

Statistical rigor extended beyond single-run results. The team repeated training with three random seeds, 42, 123, and 2026, keeping the same patient-wise split and protocol, and summarized run-to-run variability with mean and standard deviation. Patient-level metrics were averaged across runs, 95 percent confidence intervals were estimated with 10,000 nonparametric bootstrap resamples of the 56 test patients, and paired two-sided Wilcoxon signed-rank tests with Holm correction assessed model differences. ResNet50-U-Net differed significantly from CNN for all three subregions, with Holm-adjusted p-values of 9.61 times 10 to the minus 6, 1.01 times 10 to the minus 5, and 1.17 times 10 to the minus 7. However, its advantage over U-Net was statistically significant only for Whole Tumor, with an adjusted p-value of 0.0063, and not for Tumor Core or Enhancing Tumor, where the adjusted p-values were 0.0796 and 0.6596. The authors explicitly note that the statistical comparison does not show uniform superiority of ResNet50-U-Net over U-Net across all tumor subregions.

Ablation analysis, conducted under a separate reduced-computation protocol, isolated the contribution of individual design components. Removing ImageNet initialization reduced Dice scores from 0.8847, 0.7418, and 0.6660 for WT, TC, and ET to 0.8406, 0.6794, and 0.5839 respectively, while removing encoder-decoder skip connections produced the largest performance reduction, particularly for Enhancing Tumor. Using a simplified BCE-Dice loss alone also degraded performance across all three regions. Intriguingly, replacing the learnable modality adapter with a fixed three-channel mapping slightly reduced Whole Tumor Dice but improved TC and ET Dice, leading the authors to conclude that the adapter’s effect is region-dependent and should not be interpreted as uniformly beneficial.

Perhaps the most consequential test was cross-dataset generalization. Because the original BraTS2020 checkpoint was unavailable, the team independently retrained an instance of the same architecture, first establishing an internal reference on the same 56-patient BraTS2020 test cohort, where it achieved Dice scores of 0.8903, 0.7609, and 0.8087 for WT, TC, and ET. That retrained checkpoint was then evaluated on 95 adult diffuse glioma cases from the BraTS-Africa challenge cohort, with no fine-tuning, parameter updating, or threshold optimization using African data. Patient-level Dice scores of 0.8592 plus or minus 0.1674, 0.7387 plus or minus 0.2500, and 0.6990 plus or minus 0.2310 were obtained for WT, TC, and ET respectively, with corresponding HD95 values of 13.19, 7.24, and 7.03. The results demonstrate that the framework retains meaningful segmentation capability under dataset shift, though robustness varies across subregions and inter-patient variability increases for TC and ET.

The authors are candid about limitations. The observed metric improvements do not establish direct clinical benefit, since outputs were not prospectively assessed by clinical experts, and further clinical validation is required before practical use. The statistical analysis was limited to three runs and a single fixed patient-wise split, and the framework performs 2D slice-based segmentation rather than full 3D volumetric prediction, meaning its HD95 values should not be compared directly to official 3D BraTS leaderboard figures. Evaluation times, measured on a desktop system with an Intel Core i9-12900K processor and NVIDIA GeForce GTX GPU, ranged from 9.41 seconds per patient for the CNN to 45.46 seconds for ResNet50-U-Net, reflecting the cost of the deeper residual encoder. Qualitative examination of challenging cases revealed failure modes including missed small enhancing components, under-segmentation, and boundary mismatch, with CNN or U-Net occasionally providing more accurate delineation than the proposed model. The researchers point to full 3D volumetric evaluation, additional multi-institutional datasets, and prospective clinical evaluation as the necessary next steps for this line of work.

Subject of Research: Deep learning segmentation of brain tumor subregions from multimodal MRI using a ResNet50-U-Net framework

Article Title: ResNet50-U-Net integrated deep residual framework for brain tumor subregion segmentation using multimodal MRI

Article References: Akkineni, H., Bala, M. M., Rao, D. M., & Srinivasu, P. N. (2026). ResNet50-U-Net integrated deep residual framework for brain tumor subregion segmentation using multimodal MRI. Discover Artificial Intelligence, 6(1), Article 1592. https://doi.org/10.1007/s44163-026-02548-7

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02548-7

Keywords: brain tumor segmentation, ResNet50, U-Net, multimodal MRI, BraTS2020, BraTS-Africa, deep learning, glioma, Dice similarity coefficient, encoder-decoder, medical imaging, cross-dataset validation