T-CAIREM Trainee Rounds: Dilakshan Srikanthan & Ariana Walji

Join us for these presentations

Trainee Rounds Presentations (Session 4)

Dilakshan Srikanthan

Dilakshan Srikanthan is an MD/PhD candidate in Translational Medicine at Queen's University, supervised by Dr. Parvin Mousavi and Dr. John Rudan. His doctoral research focuses on applying deep learning to glioblastoma detection, characterization, and prognostication across the clinical care continuum. His work spans two domains: intraoperative mass spectrometry, where he developed cross-cancer self-supervised learning and uncertainty estimation methods for real-time tumour margin detection using REIMS, and the applications of foundation models in computational pathology for glioblastoma characterization and prognostication. Before Queen's, he trained at the Hospital for Sick Children and University of Toronto under Drs. James Rutka and Cynthia Hawkins, where he investigated immunotherapeutic approaches to diffuse midline gliomas during his Master's. Dilakshan is a recipient of the CIHR Vanier Scholarship, and the Ontario Graduate Scholarship. Aside from this, Dilakshan enjoys playing soccer, running, training in Muay Thai, and spending time with family and friends. 

Abstract title

Spatial Interfaces Between Histological Regions Predict Glioblastoma Prognosis Beyond Composition

Abstract

Glioblastoma (GBM) is characterized by marked spatial heterogeneity, yet computational pathology studies have largely relied on region composition or uninterpretable deep features to predict survival. Here, we tested the hypothesis that the spatial arrangement of histologically defined tumor regions carries prognostic information beyond composition alone. We classified 1,032 whole-slide H&E images from 416 patients across CPTAC and TCGA into seven IvyGAP-defined anatomical regions using six encoders, including five pathology foundation models and an ImageNet-pretrained ResNet50 baseline. From patch-level predictions, we constructed k-nearest-neighbor spatial graphs and derived region proportions and 28 spatial architecture features per slide. Survival associations were evaluated using three-tier Cox regression (univariate, age/sex/cohort-adjusted, and fully adjusted for region proportions), with cross-encoder replication used as an internal validation strategy. Region proportions alone were weakly prognostic, whereas specific region-region interfaces showed stronger and more consistent associations with outcome. In particular, the leading edge–necrosis interface (LE-CTne) predicted shorter progression-free survival after full adjustment in 5 of 6 encoders and was directionally consistent in all 6. The infiltrating tumor–necrosis interface (IT-CTne) was the strongest feature for overall survival. Spatial transcriptomic analysis in 18 Visium samples showed that LE-CTne interface spots were enriched for mesenchymal-hypoxic, inflammatory macrophage, and immune-vascular programs, with loss of neural progenitor signatures, consistent with a pseudopalisading niche. Mutation analysis further revealed opposing spatial phenotypes for EGFR amplification and TP53 mutation, suggesting that these alterations occupy different positions along an infiltrative-versus-compact architectural axis. Together, these findings show that interpretable region-region spatial interfaces derived from routine H&E slides provide prognostic information in GBM beyond tumor composition and map to biologically coherent microenvironmental states.

Ariana Walji

Ariana (she/her) is a Research Trainee at University Health Network and an MSc candidate at the University of Toronto. Her research focuses on facilitating the clinical integration of AI decision support tools into the operating room, with an emphasis on real world usability, workflow alignment, and improving patient safety outcomes. She completed her undergraduate studies in Medical Science at Western University where she graduated Honours with Distinction.

Ariana has extensive research experience on a global scale. She previously worked as a summer research student at Princess Margaret Hospital in Toronto, contributing to machine learning projects aimed at improving radiation treatment outcomes for patients with oropharyngeal cancer and breast cancer. During her time in Amsterdam, where she completed a research minor in cardiovascular disease, Ariana led a project evaluating the clinical readiness of machine learning–based risk prediction tools for in-hospital cardiac arrest. She has travelled to Nairobi, Kenya, where she supported initiatives focused on destigmatizing mental health and improving mental health literacy in the community.

Beyond academia, Ariana is passionate about mentorship and women’s empowerment, with ongoing involvement in organizations such as Hygiene4Her and Girls Who Lead. She enjoys sustainable living and is an avid coffee enthusiast who loves trying new beans from around the world. 

Abstract title

Optimizing AI Implementation for Surgery: Recommendations for Infrastructure and Deployment in the Operating Room

Abstract

Introduction: Recent years have seen a surge in the development of Artificial Intelligence (AI) technologies to enhance intraoperative decision-making and patient safety. Yet, evidence on their real-world implementation is sparse. This study uses two well-adopted AI deployment infrastructures (Cloud and Edge computing) to investigate the impact of AI in the operating room (OR) across key implementation factors: 1) usability, 2) cost, 3) carbon footprint. Based on these findings, we aim to provide recommendations for safe, sustainable, and scalable AI implementation across OR settings.

Methods: Surgical staff (surgeons, fellows, residents, OR nurses) from a multi-site teaching hospital in Toronto, Canada engaged in Cloud and Edge system usability testing (SUS and deployment task completion rates) and provided open-ended feedback. Surgeons additionally engaged in a surgical simulation with each deployment setup to assess ergonomics (NASA-TLX and Borg-CR 10). Cost and carbon footprint were assessed based on operational metrics. Significance testing employed paired t-tests and Wilcoxon signed-rank tests.

Results: Participants (N = 23) had significantly greater deployment task completion rates with Cloud than Edge (p<0.0001), whereas SUS scores did not reveal significant differences. Significantly greater “physical” and “frustration” NASA scores were reported with the Cloud dual-monitor setup (p = 0.0004 and 0.0036 respectively), with significantly more neck exertion based on Borg CR-10 scores (p = 0.0067). Cost analysis revealed lower upfront costs with Cloud, with total costs exceeding Edge after 396 hours of use. Carbon footprint analysis revealed greater carbon emissions with Cloud (26.22 g CO2e/hour) than Edge (6.84 g  CO2e/hour). Overall end-user feedback was positive across both systems, with the highest frequency of improvement suggestions to Cloud setup and both systems’ interface.

Conclusion: This study highlights distinct trade-offs between Cloud and Edge, offering user-centred recommendations that best support the successful translation of AI into the OR.