Deep Learning Assisted Double Inversion Recovery improves Focal Lesion Detection in Optic Neuropathy.
Authors
Affiliations (2)
Affiliations (2)
- From the Department of Radiology (D.Z., Q.C., Z.L., B.S., Y.Z., X.B., Q.L., A.L.), Ophthalmology (Y.Q.), Neurology (S.W., L.C.), Qilu Hospital of Shandong University, Jinan, Shandong, China; Department of Radiology (D.Z.), The People's Hospital of Qihe, Dezhou, Shandong, China and Department of Radiology (Q.L.), The People's Hospital of XueCheng, ZaoZhuang, Shandong, China.
- From the Department of Radiology (D.Z., Q.C., Z.L., B.S., Y.Z., X.B., Q.L., A.L.), Ophthalmology (Y.Q.), Neurology (S.W., L.C.), Qilu Hospital of Shandong University, Jinan, Shandong, China; Department of Radiology (D.Z.), The People's Hospital of Qihe, Dezhou, Shandong, China and Department of Radiology (Q.L.), The People's Hospital of XueCheng, ZaoZhuang, Shandong, China. [email protected].
Abstract
To compare the acquisition time, image quality, and diagnostic confidence of DL-accelerated DIR (DIR-DL) with conventional MRI in patients undergoing imaging for suspected optic neuropathy. This prospective, study enrolled 75 consecutive patients (mean age, 37 ± 20 years; 47 female) with suspected optic neuropathy who were clinically referred for optic nerve MRI between April and September 2024. Each participant underwent both conventional MRI (including STIR, T2-FS, T1WI, and Cube T1-FS sequences) and DL-accelerated protocol (with DIR sequence in addition). Three radiologists independently assessed lesion extent, overall image quality, perceived SNR, sharpness, artifacts, and diagnostic confidence using a 4-point Likert scale (1 = poor; 4 = excellent). Lesion lengths were normalized to optic nerve length and expressed as percentages. Because of zero inflation, a two-stage approach (logistic regression for detection, linear mixed model for lesion extent) was applied. Inter-method agreement between sequence pairs was assessed using weighted Cohen's κ (quadratic weights) for segmental, total lesion length percentages and image quality scores. Interchangeability between sequences was tested using the individual equivalence index (IEI, margin 5%). Paired t-tests compared signal-to-noise ratios, and all P values were FDR-adjusted. DL reconstruction reduced total acquisition time by 43.57% (from 22 min 55 s to 12 min 56 s). DIR-DL presented a higher proportion of excellent overall image quality than conventional sequences (87.56% excellent readings vs. 82.89%). DIR and DIR-DL detected significantly longer lesion lengths than STIR (Beta = 7.59% and 7.40%, both Padj<0.001). STIR vs STIR-DL and DIR vs DIR-DL were interchangeable (IEI = 0.02 and 0.04, 90% CI within ±0.05), with almost perfect/excellent agreement (κ = 0.95 and 0.85). DIR and DIR-DL had higher specificity than STIR and STIR-DL. In a clinical setting, DIR-DL reduced scan time while providing superior image quality and comparable diagnostic confidence to standard STIR MRI for optic nerve evaluation.