The official congress poster for SERAM 2026.SERAM
Artificial intelligence once again set the tone of the scientific program. Across 300 presentations, 40 workshops and 18 parallel tracks, speakers explored AI's growing role in breast imaging, prostate MRI, oncology, neuroradiology, and emergency workflows. Yet the conversation is no longer about whether AI works, but how it can be deployed safely, effectively, and at scale.
Life-threatening emergency with unpredictable course
More than 3,000 radiologists, researchers, and imaging professionals attended SERAM 2026, which featured approximately 300 presentations, 40 workshops, and 18 parallel scientific tracks.SERAM
Deep neck infections are potentially life-threatening emergencies. Abscess formation can rapidly compromise the airway, and while many patients respond to antibiotics and drainage, others deteriorate quickly. Predicting which patients will follow a severe course remains clinically difficult.
The investigators retrospectively analyzed 178 patients. The cohort had a mean age of 48 years (range 18–93), with 104 women and 74 men. Comorbidities were common: 41% were smokers, 24.7% had hypertension, 16.3% had diabetes, and 3.9% had a cancer history.
Five algorithms, three outcomes
Clinical outcomes were substantial. Thirty-seven patients (20.8%) required ICU admission, 115 (64.6%) underwent surgical drainage, and 22 (12.4%) needed tracheostomy.
Median inflammatory markers at admission were markedly elevated, leukocytes 15,000/µL, neutrophils 12,310/µL, and C-reactive protein (CRP) 125.5 mg/L. CT analysis revealed a median abscess volume of 6.13 cm³ and a median lateral airway diameter of 10 mm, with a median of three cervical spaces showing inflammatory edema.
The team evaluated four machine-learning algorithms -- Random Forest, Support Vector Machine, AdaBoost, Gradient Boosting and Logistic Regression -- against three clinical outcomes, using stratified k-fold cross-validation and a held-out test set.
100% sensitivity for ICU triage
For predicting ICU admission, AdaBoost performed best, achieving an AUC of 0.959 with 100% sensitivity -- meaning the model correctly flagged every patient who ultimately required intensive care. Accuracy was 86.4%, and the most influential variables were CRP, neutrophils, and creatinine.
An AdaBoost model identified all ICU-bound patients in a cohort of 178 deep neck infection cases, achieving 100% sensitivity.SERAM
Tracheostomy prediction was more challenging but still clinically meaningful. An ensemble model weighted primarily toward logistic regression achieved 83.3% sensitivity and 74.4% specificity, with an AUC of 0.799. The high negative predictive value (96.7%) suggests the model may be particularly useful for ruling out tracheostomy risk.
Key predictors and bedside thresholds
Predicting the exact type of surgical drainage proved the hardest task. The best individual model reached an accuracy of 57.8%, with the ensemble marginally lower at 55.6%, though specificity of 79.4% indicated reasonable capacity to identify patients who would not require surgery.
Key risk thresholds for severe deep neck infections derived from machine-learning analysis.SERAM
Beyond the machine-learning outputs, the researchers derived simplified decision thresholds from decision tree analysis, designed for potential bedside use. For ICU admission, a CRP above approximately 198 mg/L, neutrophil count above 17,200/µL, and abscess volume above 5 cm³ flagged higher risk. For surgical intervention, a CRP threshold of around 182 mg/L emerged as the strongest single predictor. For tracheostomy, airway diameters at or below 9 mm on the lateral axis were associated with significantly elevated risk across all three outcomes.
The authors concluded that CRP, neutrophil counts, abscess size and airway stenosis are the dominant indicators of disease severity, and that AI models, particularly for ICU triage, show promising performance.
They called for external multicenter validation before clinical implementation, and suggested that a decision-support application could eventually integrate these thresholds into routine workflow.
