A multimodal attention-augmented deep learning framework for oral cancer diagnosis, prognosis and personalised treatment recommendation

Despite advances in the diagnosis and prognosis of oral cancer, survival remains poor and early diagnosis is difficult owing to disease heterogeneity and late presentation. Single-modality AI systems based on clinical photographs, histopathology or radiology offer limited predictive performance and do not support downstream treatment decisions. This study proposes XceptionAttnV1, a multimodal attention-augmented framework that jointly processes clinical metadata, imaging (histology, CT, MRI), spectral measurements (Raman, near-infrared) and genomic features through dedicated feature extractors, a cross-modal attention fusion layer, and four prediction heads: a lesion classifier, a DeepSurv-style overall-survival head, a multi-task recurrence/lymph-node-metastasis (LNM) head, and a hybrid deep-learning-plus-rule-based treatment-recommendation head. The imaging branch couples an Xception backbone with the Convolutional Block Attention Module (CBAM); the spectral branch is a 1-D convolutional network with an LSTM tail; the genomic branch is a variational autoencoder; and only the final projection layer of the fusion module is Bayesian (a mean-field variational last layer) for epistemic-uncertainty estimation. The primary clinical scope is oral cavity squamous cell carcinoma (OSCC) and oral potentially malignant disorders; where pooled sources also contain oropharyngeal or laryngeal HNSC, those records are analysed as a labelled sensitivity subset. Because no public archive supplies all four modalities for one oral-cancer cohort, each endpoint is evaluated only on the sources that carry the relevant modality and label, and the modality composition of every evaluated partition is declared explicitly; the reported figures are therefore endpoint-specific rather than a single four-modality benchmark. This study states plainly at the outset: no endpoint in this study evaluates all four modalities together, and the framework should not be read as a validated four-modality system—it is a shared architecture evaluated separately, and honestly, on the modality-specific partitions that the available data actually support. On these task-specific held-out partitions XceptionAttnV1 attains an AUC of 0.961 for lesion classification (an imaging-plus-clinical endpoint evaluated on oral-cavity histology cohorts), and, on the matched TCGA-HNSC subset for which clinical, imaging and genomic data are co-registered, a C-index of 0.781 for overall survival, an AUC of 0.892 for LNM, and 87.4% top-1 concordance with multidisciplinary tumour-board decisions, exceeding strong single-modality and conventional multimodal baselines evaluated under the same modality availability. Ablations confirm that cross-modal attention and the Bayesian last layer contribute separately to accuracy and calibrated uncertainty. The framework is developed through a six-phase methodology; Phase 6 (prospective clinical validation) is defined as future work and has not been carried out, and all results reported here are retrospective. Consequently, the framework is presented as a retrospective research prototype and structured second opinion, not as a system validated or ready for clinical deployment.