How can we better predict if a child’s brain tumor will return—before symptoms start?. In our new study @NEJM_AI, we used deep learning to analyze longitudinal brain MRIs—learning how scans change over time—to predict tumor recurrence in children. Let’s dive in. 🧠📊 👇.🧵1/

RT @manorlaboratory: My lab’s 5-year NIH R01 grant, awarded to study gene therapy for hearing loss, was abruptly terminated. I want to shar….

RT @yekeduz_emre: Lean body mass–based oxaliplatin dosing reduces peripheral neuropathy in stage III colon cancer without compromising surv….

RT @DrUppaluri: Here it is! KEYNOTE-689 primary publication online now @NEJM. It’s been an incredible journey with DAdkins @SitemanCenter @….

It was a real honor working with @anantm and @CJTsaiMDPhD on this companion piece to our ASCO education session. Exciting things ahead for AI applied to #HNSCC clinical care.

RT @jryckman3: 🧵1/. Big news in unresectable stage III NSCLC. The InTRist study is the first randomized trial comparing NAC vs. neoadjuvan….

#ASCO25 #ASCO.

Very interesting and thought-provoking discussion. Thank you to my co-panelists and all in the audience for joining. @ASCO #ASCO2025.

RT @MGBResearchNews: In a new study, researchers from @BrighamRadOnc and colleagues used an #AI tool to help predict relapse of pediatric b….

RT @BenjaminKannMD: 11/.The model and code are open-source 🔗. 🤖Code/Model: 📰Paper (free access link): https://t.c….

12/ .This work was a multi-institutional effort led by 🌟 Phd student @tak_divyanshu. A huge thank you to all collaborators. @MGBResearchNews @DanaFarberNews @BostonChildrens @wearecbtn @aim_harvard . #MedTwitter #AI #Pediatrics #NeuroOncology.

11/.The model and code are open-source 🔗. 🤖Code/Model: 📰Paper (free access link): 🗞️Press release:

10/.This is the first self-supervised deep learning framework designed for longitudinal brain tumor imaging in kids. The temporal learning framework can be applied to any disease where serial imaging is used—cancer, chronic illness, stroke, etc. We hope to see this explored!.

9/. Why it matters:. ✅ High-risk patients could initiate early, timely treatment before symptoms . ✅ Low-risk patients could be put on less frequent scan surveillance.

8/. What does this mean? We can now analyze a patient's historical MRI data and provide a point-of-care prediction of short-term recurrence, i.e. individualized recurrence prediction.

7/. More scans = better predictions. Model performance improved the more MRI time points it had, usually stabilizing around 3–6 prior scans, and this varied depending on the dataset.

6/.Once the model can determine scan order, then comes the real test:. Can the model be fine-tuned to predict if a glioma will recur within 1 year?. Yes—and impressively. It beat standard models by up to 58% in F1-score across 3 datasets of 715 children and 3,994 MRIs. 📈

5/. The model is first trained to do a pretext task: simply learn to guess whether MRI scans are shown in the right chronological order. This self-supervised trick helps the model learn to focus on temporal patterns within the tumor region—without needing labeled outcomes.

4/. When we as clinicians analyze brain tumor scans, we explicitly compare the patient's current scan to their past scans. How can we teach AI to do the same?.

3/. Enter: temporal deep learning. Inspired by how self-driving cars learn from video data, temporal learning teaches itself the order of MRI scans to understand how tumors evolve over time.

2/. Pediatric gliomas vary wildly in how and when they come back after surgery. But we treat all patients with the same long-term MRI surveillance—costly, stressful, not tailored. We need personalized prediction of recurrence to decide who needs more therapy (or) fewer MRIs.