We used publicly available Magnetic Resonance Imaging (MRI) from The Cancer Genome Atlas (TCGA) (Clark et al., 2013). Glioblastoma Multiforme (GBM) and Low Grade Glioma (LGG) collection (Bakas et al., 2017a,b). This included the skull-stripped and co-registered MICCAI-BraTS 2018 Test Dataset (Menze et al., 2015; Bakas et al., 2017c). The data consisted of pre-operative multimodal MR imaging sequences (i.e., T1, T1-Gd, T2, T2-FLAIR) along with their whole-tumor segmentation labels composed of edema, enhancing tumor, and non-enhancing tumor. We combined these labels into a single whole tumor for this study. Number of patients in GBM BraTS Test Dataset and LGG BraTS Test Dataset were split approximately in half and allotted to validation and test datasets. The GBM and LGG data were merged across the three categories to form an ALL dataset. Each patient was associated with 144 pre-operative MRI scans, which were treated as independent data points for 2D segmentation. These MRI scans were cropped to 144 × 144 pixels and further pre-processed the data by pixel-intensity normalization. Table 1 describes the total number of patients and total number of MRI scans available in each dataset. The training datasets were used for model development (section 2.2), validation datasets were used to determine hyperparameters and calibrate the models (section 2.3), and test datasets (D_GBM, D_LGG, D_ALL) were used to perform subsequent experiments (section 3).

We used the state-of-the-art U-Net architecture (Ronneberger et al., 2015) to develop three tumor segmentation models using the GBM, LGG, and ALL train datasets. The U-Net architecture consists of an encoder, decoder, and skip connections. Each module of the encoder consists of 2D Convolution layers, followed by Batch Normalization and MaxPooling layers. Four such modules make up the encoder. The decoder consists of four modules of Conv2DTranspose layers followed by Concatenate layers. The network performs slice-wise (2D) segmentation with multi-modal MRI scans provided as the input. Models were trained with Dice Loss function for 100 epochs on 8 GPUs. Adam optimizer (Kingma, 2015) was used with a learning rate of 1 × 10⁻⁴ and a batch size of 128. Data augmentation was used while training each of the models to improve generalization. This consisted of random rotations (0–25° degrees range), random zooming (value = 0.2, zooms image by 80–120% range), width shift (value = 0.2, horizontal translation of images by 0.2 percent), height shift (value = 0.2, vertical translation of images by 0.2 percent), shear (value = 0.2, clips the image in counter-clockwise direction) and random horizontal flips. We referred to publicly available code for model development, model training, and data augmentation (Dong et al., 2017; Ojika et al., 2020).

The goal of model calibration is to align the algorithm's predicted probabilities align with the observed (ground truth) outcomes (Guo et al., 2017). Calibration process ensure that algorithms do not overstate or understate their confidence in prediction of tumor (Jungo and Reyes, 2019; Mehrtash et al., 2020). Models that have been already trained can be calibrated with post-hoc methods (Rousseau et al., 2021). Guo et al. (2017) recommend performing post-hoc calibration with the same validation dataset (Table 1) used for model development. We use Platt Scaling technique (Platt, 1999) for post-hoc calibration due to its simplicity and ease of implementation. To ensure models are properly calibrated, we compute and report common calibration metrics. Negative Log Likelihood (NLL) measures a probabilistic model's quality and is also known as cross-entropy loss. Brier Score (BS) measures the accuracy of probabilistic predictors. Percentage Expected Calibration Error (ECE%) partitions the model's predictions into equally spaced bins and takes a weighted average of the difference between accuracy and model confidence across bins. Percentage maximum calibration error (MCE%) estimates the worst-case deviation between confidence and accuracy. For metric definitions and more information, we refer readers to Mehrtash et al. (2020) and Guo et al. (2017). Table 2 indicates that all models are properly calibrated.

There exist a plethora of metrics to evaluate the performance of a medical image segmentation algorithm (Udupa et al., 2006; Taha and Hanbury, 2015). Each metric focuses on a specific aspect of the algorithm's performance, and is thus limited in capability to describe the algorithm's performance by itself. Several metrics are necessary to describe comprehensive characteristics of segmentation performance (Renard et al., 2020).

We perform this experiment as a baseline, reflective of the current standard practice for evaluation. We follow the guidelines described by Taha and Hanbury (2015) and select eight metrics to evaluate segmentation performance. Sensitivity (Sens) measures the proportion of tumor pixels that are correctly identified as tumor (foreground). Specificity (Spec) measures the proportion of benign pixels that are correctly identified as benign (background). Positive Predictive Value (PPV) measures the probability that pixels classified as benign truly belong to parts of the patients' brain without a tumor. Negative Predictive Value (NPV) measures the probability that pixels classified as tumor truly belong to parts of the patients' brain with a tumor. While accuracy can be skewed due to the paucity of tumor pixels in the tumor class, Balanced Accuracy (BAcc) takes into account class imbalance. Dice Coefficient (Dice) and Jaccard Coefficient (Jac.C) both measure the overlap between tumor annotated by the different sources (ML algorithm and the radiologists' manual annotations). Area under Receiver Operating Characteristics curve (AUROC) describes the probability that a randomly selected tumor pixel will have a higher predicted probability of being a tumor than a randomly selected benign pixel. We eliminate any extra-cranial regions and only consider the skull-stripped brain for computing the metrics. We compute metrics on a per-patient level, as it offers more granularity than at a population-level.

Prediction Intervals (PIs) are often reported and considered for medical decision-making (Kümmel et al., 2018). In radiation oncology, Chan et al. (2008) used prediction intervals to capture uncertainty in tumor and organ movement. While a confidence interval measures the precision of a predicted value, PIs measure the expected range where a future observation would fall, given what has already been observed. The width of the PI is directly proportional to the model uncertainty at that region (Kabir et al., 2018). We use prediction intervals to quantify uncertainty in tumor segmentation.

We use Conformal Quantile Regression (CQR) (Romano et al., 2019) to compute PIs. Construction of PIs is difficult, as PIs can be too small that they don't capture the true magnitude (Type 1 error) or too large that they are uninformative (Type 2 error) (Elder et al., 2021). The CQR method guarantees construction of PI such that the target value is contained within the PI by error probability α (valid coverage) and that the PIs are informative.

We used the CQR method to compute PIs in a post-hoc manner. The method uses a dataset for training the CQR models and a separate test dataset to compute the PIs. To reduce computational cost, we selected summary images (image with the largest tumor) for each patient in the validation and test datasets (Table 1). We designed a setup to generate prediction intervals around the calibrated model values. We first obtained logits (model output before the calibration) for the selected summary images for patients in both datasets. The CQR models were trained on validation dataset logits and the corresponding calibrated model predictions as target values. The trained CQR models were then used to compute prediction intervals for test dataset logits. We followed the method described by Romano et al. (2019) to compute average prediction intervals (API) per-patient in the test set. We then generated API box plots for all nine conditions.

This experiment was designed to test the impact of data quality and potential batch effects on the predictions of the model. There has been a lot of work in other domains on evaluating the adversarial robustness of ML algorithms. The application of imperceptible noise can change the prediction of image classification system from correctly identifying a panda to confidently miscalling the image a gibbon (Goodfellow et al., 2015). There are now a variety of adversarial attack techniques, from white-box techniques that can look inside the algorithm to those that can build attacks simply by testing inputs and outputs. These techniques can provide a useful framework for evaluating the robustness of a medical imaging machine learning system. In tumor imaging in general, Zwanenburg et al. (2019) showed how radiomics features can be evaluated for robustness by perturbing the tumor mask. Understanding how vulnerable ML algorithms are to noise, and how easily they change their decisions in response, gives a sense of how these ML algorithms might fail.

The adversarial attack used in this experiment was fast gradient signed method (FGSM), described by Goodfellow et al. (2015). This technique is a white-box method which takes the calculated gradient of the neural network to find the direction of the smallest change that will affect the label of the output. This gradient adversarial noise is multiplied by a factor of epsilon, to vary the strength of the attack. In these experiments the epsilon factor was varied over a range of 0–1 (0, 0.005, 0.01, 0.05, 0.1, 0.2, 0.4, 0.6, 0.8, 1.0), with more examples on the lower end of the range to evaluate small perturbations.

We performed the FGSM attack on each of the test datasets (D_GBM, D_LGG, and D_ALL), for all three ML algorithms (M_GBM, M_LGG, and M_ALL). The full panel of metrics was computed for each of these experiments. The performance of the ML algorithms was expected to decay as epsilon decreased, but the relative robustness of each of the ML algorithms and the way that they decayed was studied as well. The chosen epsilon values were (0, 0.005, 0.01, 0.05, 0.1, 0.2, 0.4, 0.6, 0.8, and 1). An epsilon of 0 indicates that no change was made to the image.

Magnetic resonance imaging has many strengths in studying and monitoring cancer status, including a variety of sequences to investigate different aspects of tumors. However, the flexibility it provides to radiologists can lead to inconsistencies in protocol and scan quality. MRQy is MRI quality package that provides a variety of features that assess the quality of a scan, and other effects that might be considered batch effects (Sadri et al., 2020). The complexity of machine learning algorithms makes it possible for them to pick up on batch effects between sites rather than the underlying biology of a problem.

These MRQy factors were used to audit the susceptibility of the different ML algorithms to scan quality factors. For each of the MRI sequences, MRQy features were calculated independently on the original NIFTI files. The features used per modality were: MEAN, RNG, VAR, CV, PSNR, SNR1, SNR2, SNR3, SNR4, CNR, CVP, CJV, EFC, TSNEX, TSNEY, UMAPX, UMAPY (For metric definitions, Sadri et al., 2020). The metadata and size features were excluded as they were not available, and the sizing was consistent across all the images. The average true positive probability of a tumor pixel having a tumor label was calculated, as well as for true negative, false positive and false negative pixels. These were calculated on a per patient level and then averaged across all the patients in the test set. These values along with Dice score and AUROC were then assessed for their correlation with the MRQy features using Spearman correlation coefficient. MRQy features that are correlated with model performance are potential quality control metrics that might be used to flag problematic cases. False discovery rate (FDR) correction was then performed using Benjamini-Hochberg correction at an alpha of 0.25 (Benjamini and Hochberg, 1995). We used this correction as it is less stringent than a more aggressive Bonferroni correction and was still found to eliminate the uncorrected p-values.

Additionally, independent of the metrics, batch effects were investigated using the MRQy parameters to compare TCGA site codes in the combined testing data set (D_ALL). The MRQy features were normalized then decomposed using principal component analysis (Tipping and Bishop, 1999). The first two MRQy principal components and their relationship to institution were investigated using ANOVA and paired T-tests in the statsmodels python package (Seabold and Perktold, 2010). We hypothesized that some site differences within the data sets might be captured by this dimensionality reduction.

True Negative (TN) panel in Figure 2 indicates all models perform equally well in identifying benign pixels. M_ALL has the highest percentage TP, indicating the best performance at correctly identifying tumor pixels. On average, due to a higher percentage of False Negatives than False Positives, all algorithms (M_LGG,M_GBM,M_ALL) under-segment tumor more often than they over-segment. The FP value is highest for M_LGG. Thus, out of all models, M_LGG classifies benign regions as tumor the most (over-segments). The FN value is highest for M_GBM, on average. M_GBM thus, under-estimates tumor pixels and classifies them as benign (under-segments).

The training of the algorithms further explains these findings. M_LGG learns to pick up subtle patterns in the training phase, and when evaluated on D_GBM, classifies normal-appearing tissue as part of a tumor. In contrast, M_GBM is used to seeing dominant contrast patterns, which explains why it misses a lot of tumor pixels in LGG.

In Figure 3, all models have similarly worse performance on some patients, indicated by red rows. This is visible across all test datasets. This could be due to multiple confounding variables such as different vendors, field strengths, parameters of imaging, strength of the imaging magnet, type of machine, and it is difficult to pinpoint the contributing factor. Metrics show similar trends in all conditions. Models have a high specificity, low sensitivity, and a high AUROC. There is an overall trend of NPV being higher than PPV. These findings reflect the effect of class imbalance in the dataset, and the models' ability to recognize benign areas much more easily than tumor regions.

Here, we present example patients (Figures 4–7) with the Ground Truth (GT) tumor and tumor segmentation contours of M_GBM, M_LGG, and M_ALL. We selected good and bad segmentation examples from D_GBM and D_LGG each for qualitative analysis. One of the authors, who is a board-certified neuroradiologist of more than a decade of experience in brain tumor diagnosis, interpreted these images.

Violin plots were constructed to analyze average model confidence across all patients. Figure 8 depicts the average prediction intervals for the skull-stripped brain region. Models have approximately the same median average prediction intervals (API) on each test dataset. Figure 9 represents model confidence while identifying tumor regions. Models have wider inter-quartile range and greater variability compared to Figure 8. This indicates models have low confidence in identifying tumors as compared to non-tumor. M_GBM and M_ALL have similar distributions of API across patients, indicating both models are similarly confident while segmenting both GBM and LGG tumor. M_LGG has the lowest median prediction interval widths, and their distribution has the lowest variability and highest concordance. This indicates M_LGG is the most confident model while segmenting both LGG and GBM patients. Out of all models, M_LGG is consistently confident while making predictions.

M_LGG has the highest confidence, even though it makes mistakes (over-segments) in segmentation, suggestive of an aggressive approach. M_GBM also makes mistakes (under-segments) but has lower confidence, which suggests a cautious approach. LGG may be monitored for a longer period of time, so a high rate of false positives can overburden clinicians, going against the goal of reducing their burden. If mistakes are very obvious, it can cause a high degree of frustration and eventual abandonment of the algorithm (Beede et al., 2020). Previous works have proposed monitoring cases with low confidence (Kompa et al., 2021). However, in a case where a model makes mistakes with high confidence, a confidence-based screening approach might cause the reviewer to miss important areas of model failure.

The three models (M_GBM, M_LGG, M_ALL) were each evaluated on the three test datasets under FGSM attack across a range of epsilons from 0 to 1. The 95% confidence intervals are also included for each of the metrics that were evaluated on a per patient level. M_GBM was the least robust to this type of FGSM attack, across all three test datasets for AUROC, Dice score, and Sensitivity. This might be due to the somewhat consistent imaging presentation of glioblastomas. It was marginally more robust to attack on its own datatype (D_GBM). All three models failed by losing sensitivity instead of specificity, indicating that the models began drastically under-segmenting the tumor under high levels of noise. Figure 10 highlights the model behavior under different levels of noise. Under smaller amounts of noise (Figure 11), the all model had the best performance generally, though not significantly. M_LGG and M_GBM had the highest AUROC values of the three models for D_LGG and, D_GBM respectively, though the differences did not reach the significance threshold of (p < 0.05).

We found that the models trained only on D_GBM were less robust to adversarial noise, particularly at high levels of adversarial noise. These levels of noise may be extreme, but do give some sense of the performance of the models under duress. Other types of attacks that might be worthwhile to investigate include: adversarial patch attacks, Carlini and Wagner attacks, projected gradient descent, as well as GAN based attacks (Carlini and Wagner, 2017; Brown et al., 2018; Ren et al., 2020). This is not the only way of assessing robustness of models, as it assumes a motivated attacker to guide attacks, as opposed to natural sources of error, but it addresses how the margins of the tumor are affected on a consistent scale across the models. Natural sources of error are less coherent, comparable, and not as well computationally modeled in MRI as the body of work on adversarial attacks.

The calibrated models' metrics and probabilities were evaluated for correlations with MRQy parameters, across the different test datasets. While there were some limited parameters that had significant correlations with model metrics, this was before FDR correction. One Thousand two hundred and twenty-four parameter to metric comparisons (17 MRQy parameters, 4 sequences, 6 metrics, 3 models) were performed, and none of the parameter-metric pairs were significantly correlated after FDR correction (p < 0.05). The MRQy features were collected before preprocessing, and were shown to be different across different institutions. However, the model used preprocessed data, and the MRQy features were not significantly correlated with the models' predictions and performance. This negative result adds more confidence to the predictions of the machine learning pipeline.

The PCA analysis showed that there were significant differences between three groups of site codes. The first cluster of institutions was 12, 06, and 08, the second was HT, DU, CS and FG, and the last was 02. Paired t-tests showed that the first principal component created splits with significant differences (p < 0.05). Notably, the numerical codes (02, 06, 08, 12) correspond to GBM studies, and alpha codes corresponded to LGG studies (HT, DU, CS, FG). However, within these clusters, the differences didn't reach significance. Figure 12 shows the site codes plotted in PCA space, and then the three models with Dice coefficient. The fact that Henry Ford Hospital (06 for GBM and DU for LGG) had more in common with other GBM and LGG sites than between those two groups is notable, though hard to explain with such a limited sample size. Site 02 was also an outlier from both other clusters in this PCA space, and had relatively poor performance, though with one case it's hard to draw a firm conclusion.

The BraTS 2018 test datasets (Menze et al., 2015) did not have significant correlations after FDR correction between scan quality and metrics. This could be due to the high-fidelity curation and good consistency of the dataset. Another potential explanation could be the limited size of the dataset. Still, these data quality metrics show significant correlations with TCGA sites after PCA analysis, indicating batch effect differences, at least between the GBM and LGG datasets. Other data quality issues that models should be tested for include bias based on race, sex, and socioeconomic status. The rise of federated learning models makes this more urgent, because they allow for training models across collaborators without sharing data (Kairouz et al., 2021). Since sensitive data is not shared between sites, tracking batch effects and sources of bias requires more work and planning than if all the data were shared and managed centrally.

Despite the technological advancements of Machine Learning (ML), the adoption of Ml in clinical workflows remains limited (Caruana et al., 2015; Strickland, 2019; Beede et al., 2020). This divide between the development and adoption of ML algorithms has been termed the “translation gap” (Steiner et al., 2021). This limitation is in part due to lack of holistic evaluation of the performance of those ML systems.

Majority of existing algorithms are statistically validated only using segmentation metrics (van Kempen et al., 2021), such as Dice Coefficient (Dice, 1945). In our experiments, we followed guidelines (Taha and Hanbury, 2015) to compute several segmentation metrics and test the differences between segmentation of GBM and LGG patients. We hypothesized that segmentation of LGG patients would be more difficult than GBM patients. LGG is diffuse and has low proliferation, which makes accurate segmentation of submicroscopic tumor tissues and tendrils, a difficult task. In contrast, GBM has greater signal intensity and characteristic presence of necrotic cavities, which makes segmentation comparatively more obvious. Our results found that metrics alone were insufficient to highlight the severity of mistakes that models make in segmentation. Only when segmentation contours were interpreted by a board-certified neuroradiologist, the degree, and types of errors of these models were evident. Similarly, in a recent systematic review of glioma segmentation algorithms, van Kempen et al. (2021) expected to find performance differences in segmentation of HGGs and LGGs but found that reported metrics could not capture such differences.

This points to a bigger concern raised by Reinke et al. (2021) that metrics alone are insufficient to evaluate all aspects of segmentation performance. While metrics are important for objective performance evaluation, they have several limitations for clinical utility (Maier-Hein et al., 2018). Difference in consequences of an algorithm's errors cannot be uncovered by metrics alone, and requires a clinical expert to elucidate them. For example, the consequences of under-segmenting in D_GBM might be more severe than under-segmenting in D_LGG due to the prognosis and management of the two diseases. As LGGs may merit a more conservatory, “wait-and-watch” approach, tumor that might be previously missed can be caught with additional tests. However, segmentation in case of GBM has more immediate consequences for resection and radiotherapy. Under-segmentation in this case would result in non-total resection, and perhaps if tumor tissue remains, would increase the likelihood of recurrence. Over-segmentation on the other hand would cause removal of non-tumor regions of the brain, or subject them to higher levels of radiotherapy, potentially causing functional impairments for patients. In case of glioma, the Dice Coefficient has a limited utility for evaluation of multifocal lesions (Giannopoulos and Kyritsis, 2010) because it cannot represent over-segmentation and under-segmentation (Yeghiazaryan and Voiculescu, 2018), does not support segmentation of multiple structures (Yeghiazaryan and Voiculescu, 2018), and is not immune to imaging artifacts and shape differences (Reinke et al., 2021). This serves as a cautionary tale that metrics alone are insufficient for reporting model performance, and there is clearly a need for better evaluation and reporting standards (Nagendran et al., 2020).

Since medical data is tightly controlled to protect patient privacy, federated learning has risen as a methodology to train models without exposing data. However, while the cross-site training structure has it's advantages, it requires thoughtful planning of model evaluation since model designers will not have access to the underlying data from other sites. Any metrics, quality control features, and batch effect monitoring will have to be carefully pre-planned to judge any resulting models. Thorough and holistic evaluation is especially important as site variability in protocol and patient populations is a known confounding factor. Our framework also helps illuminate the axes on which a federated learning network should judge their models beyond simple metrics like accuracy or AUROC.

The goal of our work is to inform how researchers can holistically evaluate their segmentation algorithms, and consider other axes of model performance than metrics alone. A problem faced by model developers in this domain is the lack of large datasets to effectively train and evaluate their algorithms. To realistically recreate this, we worked with smaller test datasets from TCGA-GBM and TCGA-LGG. Our work explores the effects of working with limited data, and informs how to interpret results meaningfully in such scenarios. Our experiments and methodology stand independently of whether the model evaluator has pre-built models, or is yet to train them. Our framework considers tumor heterogeneity, limitations of metrics and evaluates other axes such as model confidence, robustness, and batch effects. We don't suggest completely abandoning metrics—they would be important as a start, to get some level of insight. However, we caution against solely relying on metrics, and propose a more holistic evaluation of algorithms. In Figure 1, we map the axes of evaluation onto the standard ML pipeline. We provide other potential experiments that researchers can choose for model evaluation along specific axes. For example, techniques such as model ensembles and k-fold cross validation can be used to evaluate model confidence.

In our experiments, we evaluate model robustness with adversarial attacks. Recent work has shown the importance to evaluate the models' abilities to withstand adversarial attacks, especially in high-stakes scenarios such as radiology (Wetstein et al., 2020). These attacks can arise due to strong financial interests or technical infrastructure. We designed this experiment to test how and in what way could models fail in deployment under such an attack. This could lead to appropriate safeguards being put in place. Adversarial attacks also help shed light on the decision boundary of a neural network (Woods et al., 2019), which is otherwise something of a black box. Other sources of noise could be added, but have their own complications. Adding Gaussian noise to the inputs can be difficult to calibrate and variable due to randomness. Addition of artifacts, such as motion artifacts, is complex to model, and tools for doing so are not publicly available. Further research should investigate models using these failure modes, but is outside the scope of this paper. Another axes we investigate is analyzing the dataset for batch effects. In the context of tumor segmentation, batch effects could occur when image acquisition parameters or technical variations correlate with measurement quantity (Sadri et al., 2020). This may become a major problem when it leads to incorrect conclusions (Leek et al., 2010), especially when ML algorithms learn to pick up on these patterns. Analyzing for batch effects thus becomes important, as model predictions can be correlated with confounding factors. Our experiments found that pre-processing might help in making MRI scans more homogeneous and reduce these correlations.

We demonstrated our evaluation framework on ML algorithms trained with reliable, high-fidelity. expert-annotated BraTS Datasets. To further simplify the process of model development, we used straightforward implementations such as fixed dataset split (testing/validation) and 2D segmentation to work with limited data. Model developers can certainly use more sophisticated techniques that result in higher accuracy. Despite these limitations, our experiments are aligned to the overall goal of this work. Another limitation is we consider LGG for evaluation of generalizability. While there are significant imaging differences as compared to GBM, LGG is a broad category consisting of a range of tumor types. A more clinically useful investigation would be to evaluate performance on WHO recognized genetic subtypes such as IDH-mutant vs IDH-wt or 1p/19q codeleted tumors, as the literature on tumor subtypes evolves (Louis et al., 2016). However, we defer this as future work.

Here, we summarize our work and presented the following recommendations for holistic evaluation of ML algorithms:

Accounting for tumor heterogeneity in evaluation: We focus on a specific problem of glioma, and evaluate for differences in models trained by stratification of GBM and LGG Data. The first stage in standard of care for glioma is the identification of the type, which further dictates the prognosis and treatment planning. However, there is high variability in this stage, and experts often don't reach immediate consensus. It is thus important for ML algorithms to generalize well across all tumor grades. We set out to investigate this question, by performing holistic evaluation on LGG, GBM, and mixed data. Researchers should consider unique imaging presentations of each patient and evaluate on a patient-level, as important differences might be diminished upon aggregation of data. Researchers should avoid evaluation on a dataset-level.

Adoption of tools in other domains to investigate glioma segmentation: Domains such as adversarial robustness and statistics have highly specialized tools (e.g., FGSM, conformal prediction intervals) to interrogate different aspects of model performance. In this work, we demonstrate the value of adopting such tools for the problem of performance evaluation of glioma segmentation. Our results indicate clear differences in these experiments. We found model trained on LGG Data to be more confident, and model trained on GBM to suffer the most under adversarial attacks. Researchers should evaluate their algorithms on each of the evaluation axes, by performing at least one experiment on each of the axes (Figure 1).

Exploring limitations of metrics in clinical utility: In recent years, the community has started to acknowledge the clinical limitations of standard segmentation metrics. Our work demonstrates why evaluation by metrics alone is limiting in investigating heterogeneity in clinical populations (i.e., GBM vs. LGG patients), and our findings further support recent literature. Researchers should avoid relying solely on metrics to evaluate their models.

The framework can further shed light on the practical utility of an algorithm, and serve as a decision-support tool. It is not meant to replace the triaging mechanisms already in place. Since the action that accompanies a decision is different, researchers should know the situations and the patient case before use of these algorithms. If the algorithm's prediction would be followed by a high-stakes action component such as surgery, tumor resection, or radiation therapy, accuracy of segmentation is critical. Our results indicate that algorithms trained on a specific glioma grade group do not generalize well out of distribution, so it is best to use specifically-trained models. For example, if a patient with GBM is to undergo surgery, use of M_GBM as a decision-support tool would be best. In low-stakes scenarios such as accessing the extent of tumor infiltration, generalizability is more important at the cost of accuracy. The use of M_ALL, which has knowledge of all glioma grade groups, would be best in this scenario.

Establishing a close collaboration with a clinical expert is crucial to ensure that results of the framework are appropriately interpreted. In this work, the authors collaborated with experts in neuroradiology and radiation oncology to deep-dive into the problem of brain tumor segmentation and present the limitations of metrics in a clinically meaningful way. Researchers should similarly consult a clinical expert to understand how tumor heterogeneity manifests in imaging presentations between the subgroups of the tumor they are interested to investigate. The use of this framework in other domains would thus require a close collaboration between ML researchers and clinicians for effective investigation.