Deep Learning Optimisation Strategies for Uveal Melanoma Detection Using Ultra-Widefield Pictures

Introduction

Uveal melanoma (UM) is the most typical major intraocular malignancy in adults.¹ Each extra millimetre of tumour development is related to a big improve in metastatic danger, starting from 6% to 51% relying on tumour thickness.² Therefore, early detection is important for bettering survival; nonetheless, this may be difficult as UM typically mimics benign lesions comparable to choroidal nevi or congenital hypertrophy of the retinal pigment epithelium (CHRPE). Accurate prognosis at presentation is crucial, as these circumstances carry very completely different administration plans. UM is usually handled with radiation or enucleation, whereas benign lesions are monitored for development and potential malignant transformation.^3,4

Traditional diagnostic workflows for choroidal lesions depend on multimodal imaging and medical analysis.^5–7 Modalities comparable to ultra-widefield (UWF) fundus images, ultrasonography, and optical coherence tomography are used to evaluate tumour traits and distinguish benign from malignant lesions. One downside of present apply is that picture interpretation varies extensively amongst clinicians and depends closely on particular person experience to discern advanced and delicate imaging options. These challenges are amplified in areas with restricted entry to specialised ocular oncologists, contributing to delayed prognosis, continued tumour development, and poorer prognosis.⁸ Given the aggressive nature of UM and its excessive metastatic potential, the absence of speedy, goal, and accessible triaging instruments stays a crucial hole in present care. Automated, clinician-supervised diagnostic instruments might assist tackle this want by supporting early detection and danger stratification of UM, lowering delayed intervention and over-referral.

Artificial intelligence (AI)-assisted instruments have proven potential within the detection and administration of UM. Deep studying (DL), a subset of AI, permits fashions to study patterns from information with out specific programming and is extensively utilized in medical AI analysis. Prior work on this house has investigated optimum mannequin architectures, parameters, and downstream job fine-tuning.^9–16 While these research showcase promising outcomes, mannequin efficiency can also be strongly influenced by the composition and high quality of coaching information.¹⁷ This is especially related in low-data medical settings, comparable to UM detection, the place restricted and imbalanced datasets can constrain mannequin convergence and generalisability. The impression of data-centric optimisation methods stays underexplored, and a clearer understanding of how these methods affect mannequin efficiency could assist information the event of strong instruments prepared for medical use.

This examine shifts focus from architectural decisions to enter information composition, inspecting how dataset building and enter design form DL mannequin behaviour in UWF-based UM classification. We consider the results of sophistication addition, dataset augmentation, and enhanced characteristic choice on binary classification of UM versus nevus, assessing their affect on discriminative means and calibration. By figuring out the best data-centric methods, this work establishes sensible design rules for early UM detection and presents a transferable framework for medical DL improvement in different data-limited settings.

Methods

Results

A complete of 784 distinctive sufferers recognized with UM, choroidal nevi, CHRPE, or wholesome controls have been included on this examine. The imply age and intercourse ratios have been comparable between teams. The demographic data and medical traits of the lesions are summarised in Table 1.

The classification efficiency for all fashions is summarised in Table 2. The baseline mannequin achieved an AUC rating of 0.906 ± 0.032, F1 (nevus) of 0.901 ± 0.033, and F1 (UM) of 0.816 ± 0.059. The best-performing fashions have been the wholesome class addition mannequin (AUC: 0.987 ± 0.011; F1 (nevus): 0.966 ± 0.020; F1 (UM): 0.941 ± 0.035), the ‘all-strategies’ mannequin (AUC: 0.963 ± 0.027; F1 (nevus): 0.978 ± 0.015; F1 (UM): 0.960 ± 0.026), and the ‘best-strategies’ mannequin (AUC: 0.950 ± 0.033; F1 (nevus): 0.945 ± 0.023; F1 (UM): 0.898 ± 0.046). The multi-optos mannequin was the one mannequin to show decreased efficiency in comparison with the baseline mannequin, with an AUC of 0.814 ± 0.035, F1 (nevus) of 0.877 ± 0.019, and F1 (UM) of 0.678 ± 0.049. Precision and recall for the UM and nevus teams are proven in Fig. 2. Notably, the best-strategies mannequin achieved a precision of 1.0 for UM and a recall of 1.0 for nevus, indicating zero false positives for UM and 0 missed nevi.

To consider total accuracy and calibration of probabilistic predictions, we computed the ECE. The baseline mannequin achieved an ECE of 0.130 ± 0.028. All fashions achieved superior calibration to the baseline mannequin, with ECE scores beneath 0.130. The best-calibrated fashions have been the wholesome class addition and each multi-strategy fashions, with the wholesome mannequin attaining one of the best total calibration, with an ECE of 0.049 ± 0.018. All ECE scores are illustrated in Table 2.

Discussion

DL-based approaches for early detection of UM have proven appreciable promise lately. While varied architectural designs and optimisations have been studied, data-centric optimisations stay poorly characterised regardless of potential for enchancment. In this examine, we systematically evaluated how completely different data-centric optimisation methods have an effect on DL mannequin efficiency for UM detection utilizing UWF fundus images. Several methods produced significant enhancements in discriminative efficiency and calibration, whereas others revealed essential failure modes. Together, these findings provide a principled framework to information future mannequin improvement on this and associated rare-disease classification settings.

Several methods proved notably efficient in bettering characteristic abstraction and discrimination between UM and nevi. The three highest-performing fashions have been the wholesome class addition, best-strategies (wholesome addition, handled UM inclusion, and ROI dilation), and all-strategies fashions, attaining AUC scores of 0.99, 0.95, and 0.96, respectively. The substantial efficiency and calibration positive factors noticed with the inclusion of wholesome retinal controls level to an essential precept: healthy-class grounding. By exposing the mannequin to examples of regular retinal anatomy, the coaching course of seems to anchor the mannequin’s characteristic representations in what the background ought to appear like within the absence of pathology. Without this grounding, the baseline mannequin could have partly relied on background texture and imaging artifacts as proxy alerts for sophistication discrimination. This is supported by Grad-CAM activation maps (Fig. 3). The baseline mannequin shows diffuse, nonspecific activations throughout wholesome tissue, whereas the best-strategies mannequin concentrates activation at diagnostically related areas, comparable to lesion margins and high-contrast borders. Grounding by wholesome class addition, subsequently, features as an implicit normaliser, redirecting mannequin consideration away from spurious background correlations and in the direction of true lesion-specific options.²¹

The ROI dilation experiments set up a second precept: mannequin efficiency is delicate to lesion context, and an optimum ROI dilation balances the underlying biology of the diagnostic job with out introducing an excessive amount of irrelevant information. Progressive dilation of the lesion masks revealed an optimum tradeoff between perilesional context and classification efficiency, with neither the tightest crop nor probably the most expansive area of view yielding one of the best outcomes. The best-performing configuration corresponded to a dilation through which the extra perilesional margin was equal to 50% of the unique lesion masks space, capturing diagnostically related options comparable to surrounding pigmentation, subretinal fluid, and drusen. Notably, the efficiency achieve from dilation was considerably better for F1 (UM) than for F1 (nevus). This asymmetry is probably going since UM lesions are on common bigger than nevi, which means a set proportional dilation yields a better absolute growth of the enter and exposes the mannequin to extra perilesional context. For small nevi, the identical proportional dilation provides comparatively much less contextual tissue. This sample means that DL-based UM classification has a contextual window aligned with diagnostic standards utilized by clinicians, which could be balanced to optimise efficiency, computational sources, and human annotation effort. Contextual enter ought to subsequently be handled as a tunable hyperparameter knowledgeable by medical data relatively than defaulting to full-image or arbitrary fixed-crop inputs.

Not all optimisation methods improved total efficiency. The multi-optos experiment was the one technique to lower efficiency relative to baseline (AUC 0.814 versus 0.906), illustrating that information amount doesn’t reliably translate to efficiency positive factors in medical imaging AI, an concept in keeping with present literature on language fashions.¹⁷ The addition of longitudinal photos from present sufferers elevated the overall coaching set dimension by over 250%, however launched two compounding issues. First, a number of photos from the identical affected person, even throughout timepoints, share lesion morphology, retinal anatomy, and imaging artifact profiles, growing intra-class redundancy relatively than true characteristic range. Additionally, not all extra photos met the unique high quality threshold, introducing noise that will have destabilised realized representations. In distinction, the mannequin skilled with handled photos confirmed a modest improve in efficiency, notably in F1 (UM). This is probably going because of the extra photos within the UM group. Despite the morphological modifications to the tumour from therapy, the extra information could have helped with discrimination, and the mitigated class imbalance helps clarify the superior calibration.

The addition of CHRPE as a 3rd class produced equally restricted positive factors, regardless of introducing real structural range. CHRPE lesions are characterised by sharply demarcated, densely pigmented borders that may differ considerably from nevi and UM. This means that lesion range alone is inadequate to drive significant enchancment when the added class is much less clinically or visually confusable with the goal lessons. Taken along with the multi-optos findings, these outcomes point out that probably the most precious class additions are clinically adjoining examples that assist the mannequin resolve overlapping or ambiguous options. When increasing coaching datasets, researchers ought to subsequently prioritise task-relevant range (e.g., variations in lesion morphologies, imaging circumstances, and affected person demographics) whereas sustaining applicable class stability, relatively than merely growing dataset quantity.

Beyond discrimination, calibration is a clinically crucial property of diagnostic AI instruments. In UM triaging, poor calibration carries direct medical danger, as an overconfident mannequin could assign excessive certainty to ambiguous lesions, doubtlessly suppressing applicable medical concern or referral. Across our experiments, calibration efficiency didn’t uniformly observe discrimination, with methods that improved AUC producing solely modest ECE positive factors, and others yielding enhancements in each metrics. The ROI-dilation mannequin, regardless of yielding significant positive factors in discriminative efficiency, didn’t produce a corresponding enchancment in calibration over baseline. This is in keeping with the character of the optimisation technique, as perilesional context introduces extra options enhancing discrimination however doesn’t tackle the underlying class imbalance. In distinction, wholesome class addition improved each discrimination and calibration, seemingly as a result of publicity to regular retinal anatomy gives an specific reference distribution that higher captures non-pathological variability. This reduces the tendency to assign high-confidence predictions to ambiguous or background areas, leading to chance estimates that extra precisely mirror true uncertainty. This distinction highlights that discrimination and calibration reply to completely different points of data-centric optimisation and needs to be evaluated and focused independently based mostly on the anticipated use case.

The two multi-strategy fashions collectively provide perception into how particular person optimisation positive factors work together when mixed. The best-strategies mannequin demonstrated non-additive efficiency, with efficiency just below the wholesome class addition mannequin. The all-strategies mannequin curiously confirmed marginal enhancements over the best-strategies mannequin, regardless of incorporating methods that will have elevated noise. Notably, the wholesome mannequin (AUC 0.99), which utilized a single technique, outperformed each multi-strategy mixtures, suggesting that the contribution of contextual grounding is the dominant optimisation. This discovering argues in opposition to assuming that technique stacking is at all times helpful and helps a extra selective, hypothesis-driven strategy to data-centric optimisation.

This examine has a number of limitations. This is a single-centre retrospective evaluation, and exterior validation on unbiased datasets from different establishments with completely different system producers and acquisition protocols is important to robustly assess generalisability. Importantly, the general dataset dimension stays modest relative to large-scale DL benchmarks, and the comparatively vast confidence intervals noticed throughout experiments mirror this constraint. Confidence intervals have been derived utilizing bootstrap resampling, which is understood to overestimate variance when used with small pattern sizes, underscoring the necessity for bigger, externally validated cohorts. Additionally, whereas class imbalance was assessed by calibration metrics, its affect on mannequin behaviour can’t be absolutely disentangled from different components. An essential limitation is that the bottom fact labels used have been derived from medical diagnoses on the time of preliminary presentation, that are topic to inter-physician variability; extra goal definitions, comparable to genetic profiles or disease-specific mortality, could strengthen label reliability, although these should not routinely out there. Future work ought to incorporate multicentre datasets, potential validation, and exploration of optimisations for different clinically related fashions, comparable to semantic segmentation. Integrating multimodal imaging information and extra explainability methods might additional improve medical interpretability and belief.

This examine strikes past efficiency benchmarking to supply data-centric methods to information future mannequin improvement in ophthalmic classification issues. Three rules emerge from our analyses. First, normative grounding by the inclusion of wholesome retinal controls considerably improves characteristic specificity by redirecting mannequin consideration from spurious background correlations in the direction of true lesion options. Second, contextual enter is a significant and tunable hyperparameter with an optimum lesion-to-context ratio that needs to be knowledgeable by medical data of perilesional diagnostic options. Third, information high quality and have range matter greater than sheer quantity in sure circumstances, as indiscriminate dataset growth by redundant or lower-quality photos can degrade efficiency. These rules is probably not particular to binary UM classification and will provide a transferable framework for data-centric DL improvement in rare-disease picture classification extra broadly. With multicentre validation and integration into medical workflows, fashions knowledgeable by these design rules have the potential to cut back variability in interpretation, help earlier intervention, and finally save affected person lives.

Supplementary Material

This is an inventory of supplementary recordsdata related to this preprint. Click to obtain.