1. Study Design

• Retrospective study at University Hospital Basel’s Radiology Department (Jan 2019 – Apr 2024).
• Inclusion: Adult and adolescent head CT scans (≥14 years).
• Ethics waiver granted; no external funding received.

2. Data Acquisition and Preprocessing

• Initial Dataset: 42,211 reports containing phrase “Neurokranium” reflecting internal department reporting practices.
• Pre-LLM triage to optimize workflow with phrase search: “no intracranial bleeding,” “no midline shift,” “no atrophy” / ”mild atrophy” in patients over 60 years of age: Resulting in 9,622 CT reports for further screening.
• LLM Triage :
• Locally deployed “llama3_sauerkraut:70B” (running on Ollama platform (ver. 0.2.5), temperature=0.0, internal hospital network hardware - GPU cluster with NVIDIA A100 40GB GPU).
• (1) Analysis of each radiology report for 5 categories: stroke, parenchymal defects, space-occupying lesions, enlarged ventricles, and white matter lesions beyond Fazekas 1.
• (2) Analysis of 210,761 clinical summaries from patients with normal CTs to exclude dementia, MCI, or structural brain disease impairing cognition.

Final “Normal” Cohort

• 3,086 head CTs from 2,964 unique patients.
• Age range 14–99 years; 1,633 exams in males (avg age ≈46) and 1,453 in females (avg age ≈54).

3. Deep-Learning Segmentation (nnU-Net)

• In-house nnU-Net algorithm trained on intracranial CT ventricles.
• Segmentation of the third, fourth, and lateral ventricles, subdividing the latter into frontal horns, bodies, atria (trigone), occipital horns, and temporal horns. Sample ventricular segmentation and its respective 3D reconstructions are depicted in Figure 1 and Figure 2 respectively.

Fig 1: CT imaging of a 26-year-old male patient, showing axial (a), sagittal (b), and coronal (c) slices, and their respective ventricle segmentation masks acquired through nnU-Net based algorithm (d), (e), (f).

Fig 2: Views (a–f) of the 3D reconstruction of segmentation masks shown in Figure 1 for the same 26-year-old male patient. (QR code links to the on-line interactive model).

4. Statistical Analysis

• LLM Accuracy: Estimated using 471 randomly selected radiology reports, manually labeled by a radiology resident. Sensitivity, specificity, accuracy, PPV, NPV, and F1 scores were calculated.
• Ventricular Volume: Summaries as mean ± IQR by decade age groups. Mann-Whitney U test for sex comparisons. Pearson’s r for correlations with ascending age per subcompartment.
• Tools implemented in Python 3.12.2 were used for calculations (numpy 1.26.4).