{"id":"2dcf2b8d-b3f5-4a4f-bfab-4fd899351b32","arxiv_id":"1908.04568","paper_version":3,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-headed CNN that predicts the brain midline as a probability distribution over positions and derives midline shift from the predicted curve achieves near-expert accuracy on internal and external MRI datasets.","lead":"This paper presents a deep learning method that estimates brain midline shift from MRI images by first locating the brain's midline curve and then measuring how far it bends. The method's error is close to the disagreement between human radiologists, and it keeps working on a separate routine-clinical dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'approaches inter-expert variability' claim rests on model-vs-exp1 errors compared with exp1-vs-other-expert variability, but the model is trained on exp1, so the harder test against other experts is missing.","rationale":"The paper's core technical contribution—the two-headed network with column-wise softmax and expectation decoding—is well described, the code is released, and the DS2 external evaluation is a genuine strength. The concern I raise is not about the architecture or the reproducibility of the numbers, but about the interpretation of the headline comparison. The abstract and the reader's strongest_claim both pivot on 'approaches inter-expert variability.' Because the model was optimized to reproduce exp1, the comparison in Table 2 cannot by itself establish expert-level agreement; it only shows agreement with the training expert. This is a missing control rather than a demonstrated error, so the right remedy is a conditional acceptance requiring the multi-expert evaluation or a rewording of the claim. The concrete test above would settle it. I partially agree with the reader's weakest_assumption: label reliability matters, but the more specific and actionable weakness is the asymmetry in which experts the model is compared with.","tokens_in":5712,"tokens_out":12484,"duration_ms":127923,"concrete_test":"Have two or more of the expert annotators (exp2-4) re-annotate the midline and limits on a blinded random subset of at least 30 DS1 test images (or use the existing exp2-4 annotations if they cover the test folds), then compute the proposed model's MAX and RMSE against each expert's midline and compare these values to the exp1-vs-exp2-4 rows in Table 2 using a paired bootstrap or Wilcoxon test. If the model-vs-exp2-4 errors are not significantly larger than exp1-vs-exp2-4, the inter-expert claim is confirmed; if they are significantly larger, the claim should be softened to 'matches exp1's annotations within inter-expert variability' rather than 'approaches inter-expert variability.'","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline claim of the abstract is that the method's mean error approaches inter-expert variability. The supporting evidence in Table 2 compares the model's midline errors against exp1's annotations (MAX 3.61 ± 2.62 mm, RMSE 0.79 ± 0.44 mm) with the disagreement between exp1 and exp2-4 (MAX 3.44 ± 2.13 mm, RMSE 0.77 ± 0.35 mm). Since the model was trained on exp1 labels (Section 6), this is the most favorable possible comparison: it measures how well the model mimics its training annotator, not how well it agrees with an arbitrary expert. A model that perfectly learned exp1's annotation style could score 0 against exp1 while disagreeing with exp2-4 more than exp1 disagrees with exp2-4. The paper gives no error of the model against exp2-4, no consensus or reference-standard evaluation, and no statistical assessment of the 0.17 mm MAX gap. The reader's weakest assumption about reference-label reliability is adjacent, but the specific gap is the one-sided expert comparison: the claim of 'approaching inter-expert variability' is not actually measured for the other experts.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a two-headed CNN for estimating brain midline shift (MLS) from axial MR slices. The midline head predicts, for each image row, a softmax distribution over x positions and takes its expectation as the predicted midline; a second head predicts the vertical interval (limits) where the midline is defined. The final MLS is derived from the predicted curve as the maximal deviation from the straight line connecting its endpoints. The method is trained on 352 MRI series labeled by one expert (exp1) and evaluated with 5-fold cross-validation plus an external 203-series dataset (DS2), comparing against a ResNet regression baseline and a plain segmentation variant. The paper reports MLS MAE of 1.54 mm (DS1) and 0.75 mm (DS2), a midline MAX error of 3.61 mm, and claims this approaches inter-expert variability (3.44 mm).","tokens_in":5971,"tokens_out":5075,"duration_ms":49510,"significance":"The design is principled, and the structural constraints (single midline point per row, finite limits) are derived from anatomy rather than from the test data, so the method is not circular. The softmax-expectation formulation gives a natural confidence estimate, the method is evaluated on an external clinical dataset, and the authors release training and inference code. If the reported accuracy and external validation hold under a more stringent expert comparison, the method would be a practically useful and interpretable alternative to direct MLS regression. The main caveat is that the headline 'approaches inter-expert variability' claim currently rests on a one-sided comparison with the training annotator and needs additional evaluation.","major_comments":[{"comment":"The central claim that the method's error 'approaches the inter-expert variability' is not supported by the reported comparison. The model is trained on exp1's labels (Section 6), and Table 2 compares the model's midline predictions against exp1 with the variability between exp1 and exp2-4. This measures how well the model reproduces its training annotator, not how well it agrees with an arbitrary expert; a model that perfectly imitated exp1 would score zero against exp1 while still disagreeing with exp2-4 more than exp1 does. The paper should report the model's error against exp2-4 (or against a consensus or reference standard) and, if the abstract's 'mean error' refers to MLS MAE, it should compare that metric with the corresponding expert-level MLS variability. Without these numbers the headline claim is not established.","section":"Abstract; §7.2, Table 2"},{"comment":"The proposed method's DS2 MAE is reported as 0.75 ± 0.04 mm. If the ± value is the standard deviation across the 203 DS2 series, it is implausibly small compared with DS1 (1.54 ± 1.98 mm) and with the DS2 MLS distribution (2.9 ± 1.5 mm); if it is the standard deviation across cross-validation folds or model seeds, it is not the per-case accuracy dispersion that a reader would expect. Please clarify exactly how the DS2 statistics were computed and report the per-patient standard deviation or a confidence interval.","section":"Table 1, DS2 row"},{"comment":"The limits head is a mandatory component of the pipeline because it defines the y-interval over which the midline and the MLS are computed, yet no independent metric for limits prediction is reported and no ablation removes or replaces it. An evaluation such as IoU between predicted and annotated limits, or endpoint error, would let the reader judge whether the reported gain over the ResNet baseline is due to the structural formulation or to a particular behavior of the limits head. Given that errors in limits directly propagate to MLS, this is a necessary piece of evidence for the method's internal validity.","section":"§4.2 and §7"},{"comment":"No statistical tests or confidence intervals accompany the differences that matter for the paper's claims, such as Proposed vs ResNet-152, Proposed vs segmentation, and the 3.61 mm vs 3.44 mm gap in MAX. The RMSE values in Table 2 are within 0.02 mm of the expert-variability value, so without paired significance tests, bootstrap intervals, or per-case scatter, the 'approaches inter-expert variability' claim cannot be distinguished from chance. Please add appropriate statistical assessment.","section":"§7, Tables 1 and 2"}],"minor_comments":[{"comment":"There is a typo in the phrase 'regression problem of mildine estimation'; 'mildine' should be 'midline'.","section":"Section 4"},{"comment":"The metrics MAXs and RMSEs are used in Table 2 but are not defined; please state explicitly how the per-slice maximum and root-mean-square errors are averaged over axial slices.","section":"Section 7.2"},{"comment":"The schematic would be easier to follow if the shared input layers and the two heads were labeled directly, since the text refers to 'Fig. 3 (top)' and 'Fig. 3 (bottom)'.","section":"Figure 3"},{"comment":"DS2 is described as used 'in a prospective fashion,' but the data appear to be retrospective; please rephrase to avoid ambiguity about the study design.","section":"Section 6"},{"comment":"The code link is a strength, but a specific version or commit hash would improve reproducibility.","section":"Section 1 and GitHub footnote"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead this one. It is a solid, well-scoped clinical imaging paper with a genuine methodological idea: instead of regressing midline shift directly or treating midline pixels independently, the network outputs a per-column probability distribution over x for each y, and the midline is the expected value under that distribution. A second head predicts the vertical limits where the midline is defined. It is simple, it makes sense for the task, and it is different from the ResNet regression and keypoint pipelines they cite. That is the new thing, and it is a real contribution, not a paradigm shift.\n\nWhat the paper does well: the evaluation is honest and reasonably thorough. Five-fold CV on a heterogeneous neurosurgery dataset, a prospective external dataset from routine practice, a direct ResNet baseline, and code that is public. The midline error numbers are genuinely close to expert variability, and the fact that plain segmentation does much worse on max error is a nice empirical justification for the softmax/expected-value design.\n\nNow the soft spots, in proportion. The stress-test note is half right. The abstract says the method's mean error approaches inter-expert variability, but Table 2 compares the model against exp1 only, while inter-expert variability is exp1-vs-other-experts. Since the model is trained on exp1 labels, that is the most favorable benchmark. A model that perfectly mimicked exp1 would score zero against exp1 while disagreeing with other experts as much as exp1 does. That is a legitimate caveat, and the paper should either add model-vs-exp2-4 numbers or soften the claim. That said, it is not a flaw that sinks the paper: exp1 is the experienced neuroradiologist, the practical use case is automating a measurement that exp1 would make, and the comparison to exp1-vs-others is still informative. It just is not the full inter-expert story.\n\nOther issues are minor. The DS2 MAE standard deviation of 0.04 for the proposed model looks suspiciously small next to ResNet's 1.10; the paper should clarify whether that is fold-level or patient-level spread. The limits head is never evaluated on its own, so it is hard to know how much it contributes. And there are no significance tests for the MAE/AUC differences, though the gaps look large enough that it may not matter.\n\nWho is this for? Anyone working on automated measurement in medical imaging, and specifically anyone doing landmark or curve detection with CNNs. It deserves a serious referee. I would send it to peer review and ask for the model-vs-exp2-4 numbers and a clarification of the DS2 error bar.\n\nMy take: accept with minor revisions.","headline":"A solid clinical imaging paper with a genuine methodological twist—per-column softmax midline regression plus a limits head—and a real external validation; the 'approaches inter-expert variability' claim is slightly oversold but not wrong.","tokens_in":6525,"tokens_out":1994,"would_cite":true,"duration_ms":23089,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Enforcing a per-row softmax constraint lets a two-headed CNN localize the brain midline and derive midline shift with a mean error of 1.54 mm on severe tumor cases, approaching the 3.44 mm inter-expert disagreement.","keywords":["midline shift","brain midline estimation","convolutional neural networks","softmax expectation","structural knowledge","MRI","confidence intervals","interpretability"],"falsifier":"Re-run the proposed model on DS1 with the softmax-expectation head replaced by a sigmoid+argmax segmentation head under identical preprocessing and training; if the MAX midline error does not rise materially above the reported 3.61 mm, the paper's claim that the distributional constraint is the source of the gain would be contradicted. Independently, have two additional radiologists label a random subset of DS1 and compare the model's errors against their agreement; the 'approaches inter-expert variability' claim would be falsified if the model is much farther from the new consensus than the new experts are from each other.","tokens_in":5551,"feed_emoji":"🧠","tokens_out":7711,"duration_ms":72337,"temperature":0.7,"pith_summary":"Midline shift (MLS) is a clinical marker used to predict outcomes in traumatic brain injury, stroke, and brain tumors, but automatic measurement has typically relied on multi-step keypoint pipelines. This paper argues that a convolutional network can do better when it is built around two structural facts: for each image row there is at most one midline point, and the midline exists only over a limited vertical interval. The network predicts the midline as the expected horizontal position under a row-wise softmax, then derives the shift as the maximum distance between that curve and a straight line through its endpoints. On a neurosurgery MRI dataset with many severe shifts the method reaches a mean error of 1.54 mm; on an external outpatient dataset it reaches 0.75 mm; and its worst-slice midline error (3.61 mm) comes close to the disagreement between expert radiologists (3.44 mm). If that accuracy holds, automatic MLS estimation could serve as a screening tool in settings where rapid triage matters.","feed_headline":"Softmax-constrained CNN estimates brain midline at expert-level error","feed_subtitle":"Predicting the midline as a per-row expectation cuts mean shift error to 1.54 mm on severe cases.","key_machinery":"The load-bearing object is the two-headed CNN with a row-wise softmax midline head. The head's output for each $y$ is normalized so that $\\sum_x \\mathrm{output}_{xy} = 1$, and the predicted midline position is the expected value $\\sum_x x \\cdot \\mathrm{output}_{xy}$. The second head uses global max pooling along $x$ followed by one-dimensional convolutions to predict the limits interval $\\mathcal{I}$; predictions outside $\\mathcal{I}$ are discarded, and the shift is computed as the maximum distance from the predicted curve to the straight line joining its endpoints. Because the softmax output is a probability distribution, percentiles and confidence intervals can be read off directly at inference time.","core_discovery":"The paper's central claim is that imposing a distributional constraint on the output map—$\\sum_x \\mathrm{output}_{xy}^{\\mathrm{midline}} = 1$ for each row $y$—and reading the midline as the expectation $\\sum_x x \\cdot \\mathrm{output}_{xy}^{\\mathrm{midline}}$ is what makes the difference. Under this constraint the network's job matches the anatomy: one midline point per row, with a second head deciding where the midline exists and a shared encoder for both tasks. The authors show that this softmax-expectation design outperforms both a plain segmentation head (sigmoid plus argmax) and a direct ResNet regression of the shift, and that the error is close to the inter-expert variability reported in their own data. The improvement is most visible in the maximum error across a brain, which matters because the final MLS is the maximum over axial slices.","pith_inferences":["The paper leaves implicit that the single-valued-per-row construction applies to any elongated structure with one crossing per axis; a natural next test would be the falx on CT or the spinal canal on sagittal MRI.","Because the external dataset contains mostly small shifts (mean 2.9 mm, 8% over 5 mm), the external validation mainly supports the low-shift regime; a more stressing external test would use a severe-case cohort with CT or non-routine acquisition.","A testable extension would be to replace the straight-line normal midline with a symmetry-derived reference line from the skull to see whether the residual 0.75–1.54 mm error shrinks further.","The confidence intervals from the softmax head could be turned into an automated flagging rule by thresholding their width; the paper describes the intervals but does not propose such a rule."],"forward_implications":["A network that reports MLS in near-expert agreement could be used to prioritize severe cases in settings where a radiologist is not immediately available.","Because the method works on MRI series with variable voxel spacings and modalities, it can be applied without retraining across heterogeneous hospital protocols.","The per-row probability output gives percentile-based confidence bands, so cases with uncertain midlines can be flagged for review rather than silently mis-measured.","The two-headed design removes the need for keypoint detectors, so it works even when the shift is visible only on slices without ventricles."],"supporting_citations":[{"why":"U-Net architecture provides the shared encoder-decoder backbone for both heads.","marker":"[15]"},{"why":"ResNet is both the direct-regression baseline and the source of the residual blocks used in the network.","marker":"[4]"},{"why":"The prior ResNet adaptation for head-CT critical findings defines the direct-regression baseline the authors compare against.","marker":"[2]"},{"why":"Suggests replacing plain convolutional layers with residual blocks in the segmentation backbone.","marker":"[11]"},{"why":"Reports high interrater variability in MLS estimation, framing the target of matching expert agreement.","marker":"[13]"},{"why":"Supplies the definition of midline shift and its clinical role that motivates the task.","marker":"[8]"}],"fun_headline_variants":["Softmax constraint yields expert-level brain midline shift estimates","Brain midline shift predicted at expert accuracy via softmax CNN","Task-specific CNN hits expert-level error for midline shift","Distributional constraint boosts CNN midline shift detection","Midline shift: softmax-based CNN matches expert variability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the expert-annotated midlines and limits are reliable enough to be both training targets and the reference standard for evaluation, while the paper itself reports that experts disagree by up to 3.44 mm in maximum error.","fun_headline_variants_meta":{"raw":{"variants":["Softmax constraint yields expert-level brain midline shift estimates","Brain midline shift predicted at expert accuracy via softmax CNN","Task-specific CNN hits expert-level error for midline shift","Distributional constraint boosts CNN midline shift detection","Midline shift: softmax-based CNN matches expert variability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000541,"raw_usage":{"total_tokens":2531,"prompt_tokens":821,"completion_tokens":1710,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":1634}},"tokens_in":437,"tokens_out":1710,"duration_ms":11545,"temperature":1.0,"reasoning_tokens":1634,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:38:06.691262+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the proposed model on DS1 with the softmax-expectation head replaced by a sigmoid+argmax segmentation head under identical preprocessing and training; if the MAX midline error does not rise materially above the reported 3.61 mm, the paper's claim that the distributional constraint is the source of the gain would be contradicted. Independently, have two additional radiologists label a random subset of DS1 and compare the model's errors against their agreement; the 'approaches inter-expert variability' claim would be falsified if the model is much farther from the new consensus than the new experts are from each other.","supporting_citations":[{"cited_title":"The Lancet 392(10162), 2388–2396 (2018)","cited_arxiv_id":null,"evidence_quote":"The prior ResNet adaptation for head-CT critical findings defines the direct-regression baseline the authors compare against."},{"cited_title":"Journal of Neuroimaging 28(1), 61–63 (2018)","cited_arxiv_id":null,"evidence_quote":"Reports high interrater variability in MLS estimation, framing the target of matching expert agreement."},{"cited_title":"International journal of biomedical imaging 2018 (2018)","cited_arxiv_id":null,"evidence_quote":"Supplies the definition of midline shift and its clinical role that motivates the task."}],"review_version":1}