Introduction: Cerebrospinal fluid (CSF) restricted oligoclonal bands (OCB) are the established immunological biomarker for multiple sclerosis diagnosis. Isoelectric focusing (IEF) and immunoblotting is the gold standard technique to detect OCB. Quantifying OCB remains a complex task: IEF lanes are often contaminated by band-like artefacts. Moreover, samples with low immunoglobulin G concentrations produce low-contrast lanes with faint bands hard to distinguish visually. Hence, the current analysis is subject to misinterpretations and inter-and intra-expert variabilities. We developed two automatic image-processing models to refine OCB detection in IEF. The first knowledge-driven model (OCB-ES) was designed to reproduce human decision-making processes by combining expert guidelines into relevant rules. The second data-driven model (OCB-DL) autonomously extracted relevant OCB features using a convolutional deep neural network without embedding human knowledge. Objectives: To evaluate the ability of (i) the two developed models and (ii) three experts blinded to each other’s interpretation to reproduce a visual on-membrane expert consensus analysis considered to be the ground truth. Methods: We retrospectively evaluated 982 CSF lanes from two French cohorts: POLAR and Expert-IEF. Lanes were randomly distributed into 833 for training models and optimizing algorithms’ parameter values and 149 for testing to assess performance on unseen data. At least three OCBs were required to designate a lane as oligoclonal. The performance of lane classification was evaluated in terms of sensitivity (SE) and specificity (SP). Furthermore, the two models were compared in function of the number of detected OCB using the area under the ROC curve (AUC). Results: 97 (65%) test lanes were deemed oligoclonal by the expert consensus. The three experts had the following evaluation scores: SE (0.85, 0.86, 0.84) and SP (0.79, 0.89, 0.81), respectively. OCB-ES and OCB-DL reached similar scores with SE (0.82, 0.85) and SP (0.82, 0.84) respectively. OCB-DL slightly outperformed OCB-ES however the AUC difference between OCB-DL (0.93) and OCB-SE (0.90) was not significant (p = 0.14). Conclusions: Our developed models successfully reproduced expert visual analysis and are promising easy-to-use decision support systems for OCB detection.