{"id":"edd826b1-2d69-46df-82f9-1a664566a98f","arxiv_id":"1908.11199","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Guided backpropagation gives the most accurate fine-grained localizations, and SHAP gives the most discriminative heatmaps at the dopamine-depletion site, among six interpretation methods applied to SPECT-based Parkinson's disease classifiers.","lead":"This tutorial compares six methods for explaining which brain regions a deep learning model uses when diagnosing Parkinson's disease from SPECT scans. It finds that guided backpropagation most accurately highlights the affected striatal region, while SHAP best reveals the missing uptake that separates patients from healthy controls.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The ranking of guided backpropagation as the best interpretation method rests on a striatal ground-truth mask built from fixed mean thresholds borrowed from Ref. [22]; if those thresholds mis-segment PPMI subjects, the Dice/MSE comparison and the model-selection recommendation are unsupported.","rationale":"The reader's weakest assumption is exactly the load-bearing concern I identify: the evaluation in Section III-B treats a thresholded striatal segmentation, derived from fixed mean thresholds in Ref. [22], as ground truth for interpretation quality. If that segmentation is wrong for these PPMI subjects, then the Dice-based ranking in Table IV and the related error plots are measuring overlap with an incorrect region, and the paper's main quantitative recommendation—guided backpropagation is best—is not supported. The model-selection contribution in Figure 7 inherits the same problem because it chooses Deep PD Net based on the guided backpropagation Dice score. I also agree that the metric implicitly assumes a correct interpretation must overlap the striatal uptake; a faithful model using other discriminative cues could be penalized. The reader's CONDITIONAL verdict is therefore appropriate: the claims are plausible and internally consistent, but they depend on an unvalidated ground-truth construction and a partly visual SHAP comparison. My proposed test—recomputing the evaluation with an independent, subject-specific striatal segmentation—would settle whether the concern actually changes the ranking. Since the reader already conditioned the verdict on this type of issue, no change in verdict is needed.","tokens_in":19156,"tokens_out":3508,"duration_ms":39709,"concrete_test":"Re-run the Section III-C evaluation with ground-truth masks generated per subject by a validated striatal segmentation (manual expert delineation or an atlas registered to MNI space) on the same slice-averaged images, holding all other steps (top 1% and 10% thresholds, Dice, Wilcoxon tests) fixed. If guided backpropagation is not the top-ranked method in Table IV for the four architectures, or if its margin over SHAP/DeepLIFT collapses, then the central recommendation and the Figure 7 model-selection workflow are not robust to the choice of ground truth.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The tutorial's central quantitative conclusion—that guided backpropagation is the best interpretation method—is computed against a ground-truth mask that is not independently validated for this dataset. In Section III-B the authors segment striatal nuclei by applying the mean thresholds reported in Ref. [22] (0.63 for NC, 0.69 for PD) to slice-averaged, [0,1]-normalized SPECT images, uniformly across all subjects. Ref. [22] reports those thresholds as mean ± SD of expert-selected values; using only the means discards the 0.04–0.05 SD and, more importantly, assumes that the thresholding protocol transfers exactly to this dataset and preprocessing. If the transferred thresholds over- or under-segment the striatum, then the Dice coefficients in Table IV and the MAE plots in Figure 5 are measuring overlap with an erroneous region, and the claim that guided backpropagation 'has the best ability to show fine-grained importance' is not established. A second, conceptually independent assumption is embedded in the metric: a good interpretation is defined as one whose top 1–10% pixels overlap the segmented striatum. That excludes the possibility that a faithful model uses extra-striatal cues (e.g., global asymmetry or background normalization); a model attending to those cues could be correct yet receive a low Dice score. Because the same masks drive the model-selection recommendation in Figure 7, both the main method ranking and the proposed selection workflow rest on this unvalidated ground truth.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a tutorial for selecting an interpretability method for deep convolutional neural networks (DCNNs) that classify Parkinson's disease (PD) from SPECT images. Four DCNN architectures (PD Net, Deep PD Net, and their batch-normalized variants) are trained on PPMI data, and six interpretation methods (saliency map, guided backpropagation, Grad-CAM, Guided Grad-CAM, DeepLIFT, and SHAP) are compared. The authors propose to evaluate interpretation quality by measuring Dice coefficient and mean absolute error between binarized attention maps and a striatal ground-truth segmentation derived from thresholds reported in a previous study, and they further propose using interpretation performance as a model-selection criterion. The central claims are that guided backpropagation provides the best fine-grained importance (highest Dice, lowest error) and that SHAP produces the best heatmap at the uptake depletion location, discriminating PD from normal control subjects.","tokens_in":19399,"tokens_out":3139,"duration_ms":32833,"significance":"If the central claims hold, the paper offers a reusable procedure for choosing interpretation methods in medical imaging and for using interpretation quality in model selection, with public code and data sources. The use of 10-fold cross-validation, Wilcoxon signed-rank tests, and a public dataset (PPMI) are concrete strengths that support the reproducibility of the classification and Dice comparisons. However, the significance is limited by the absence of validation for the ground-truth segmentation, the purely visual support for the SHAP claim, and the unstated construct assumption that a faithful interpretation must overlap the striatal uptake region. These issues are load-bearing because they directly support the paper's main method ranking and its proposed workflow.","major_comments":[{"comment":"The ground-truth masks used to rank interpretation methods are generated by applying the mean thresholds (0.63 for NC, 0.69 for PD) from Ref. [22] to slice-averaged, [0,1]-normalized images. These thresholds are means of expert-selected values from a different study, and the manuscript provides no evidence that the thresholding protocol transfers to this cohort and preprocessing. Because every Dice coefficient and MAE value in Table IV and Figure 5 depends on these masks, the central ranking of guided backpropagation and the model-selection recommendation in Figure 7 are not established unless the segmentation is validated. I recommend a sensitivity analysis over the reported ±0.04/±0.05 standard deviations, or comparison with an atlas-based or independently expert-derived striatal ROI, to show that the ranking is robust.","section":"Section III-B, Table IV, Figure 5"},{"comment":"The claim that SHAP outperforms other methods in discriminating PD from NC at the uptake depletion location is supported only by visual inspection of mean overlays. No quantitative metric or statistical test is reported for this claim, and it is not captured by the Dice/MAE evaluation, which actually ranks SHAP lower than several other methods. The authors should define a quantitative criterion (for example, contrast in the putamen region showing reduced uptake, or overlap with an asymmetry-based region) and report its distribution over test folds with a significance test; otherwise the second half of the central recommendation is not supported.","section":"Section IV, Figure 6"},{"comment":"The evaluation implicitly defines a faithful interpretation as one whose high-attention pixels overlap the segmented striatum. A model could be correct and faithful while relying on extra-striatal cues (for instance, global asymmetry or background normalization), and such an interpretation would receive a low Dice score. This construct-validity assumption is not stated or discussed, yet it underlies the use of the Dice coefficient as the primary evidence for guided backpropagation. The manuscript should explicitly discuss this limitation and ideally probe it, for example by evaluating whether the top-ranked method localizes class-discriminative regions that are not strictly inside the striatal mask.","section":"Section III-C, Section IV"}],"minor_comments":[{"comment":"The abstract and conclusion state that guided backpropagation has the 'lowest mean square error,' but Section III-C and Figure 5 report mean absolute error; these terms should be reconciled.","section":"Abstract, Section V"},{"comment":"The footnote says code and data sources 'will be published immediately after getting an acceptance letter,' while the introduction claims the code 'was uploaded and can be downloaded publicly'; please clarify the actual availability status.","section":"Footnote 1, Section I"},{"comment":"The text reports Wenzel et al. accuracy as 97.7%, while Table I lists 97.20%; the discrepancy should be corrected.","section":"Table I, Section II-B"},{"comment":"The description of guided backpropagation as setting the gradient of layer l to zero when the gradient of layer l+1 is negative is imprecise; the method also masks the forward activations, and this should be stated clearly.","section":"Section III-A, Subsection 1"},{"comment":"The mean absolute error plots are difficult to read in grayscale because the curves overlap; consider using distinct colors or separate panels for each method.","section":"Figure 5"}],"recommendation":"major_revision","confidential_remarks":"The central claims are plausible as tutorial-level findings, but the evidence supporting the leading recommendation is tied to an unvalidated transferred threshold, and the SHAP claim is not quantitatively supported. Both can be fixed within the manuscript's scope with additional experiments or careful reframing. I do not see grounds for rejection, but the requested revisions are more than local presentation fixes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on interpretable medical imaging. The genuinely new thing is the head-to-head comparison of six interpretation methods on SPECT-based PD recognition, with 10-fold cross-validation and Wilcoxon tests. The guided-backpropagation result is internally consistent: it wins Dice and MAE across four architectures, and the statistical testing is appropriate. The SHAP claim, by contrast, is informal — visual inspection of overlays — and the authors essentially admit it. The tutorial framing is fine; the explanation of gradient-based and additive methods is accurate, and the model-selection feedback loop (Figure 7) is a nice practical idea.\n\nThe soft spots are real and central. First, the ground truth for interpretation quality is a striatal mask made by applying mean thresholds (0.63 for NC, 0.69 for PD) from Ref. [22] uniformly to slice-averaged, [0,1]-normalized images. The original values were reported as mean ± SD (0.04–0.05), and there is no validation that the thresholding transfers to this dataset and preprocessing. If the masks are wrong, the Dice/MAE ranking and the recommendation of guided backpropagation are unsupported. The stress-test note lands. Second, the metric encodes a construct assumption: a faithful interpretation should overlap the striatum. A model attending to extra-striatal cues (e.g., global asymmetry or background normalization) could be correct yet score low. That limits what the comparison can claim; it is really a comparison of overlap with one anatomical prior, not a direct measurement of faithfulness. Third, the code availability statement is self-contradictory: the introduction says code is publicly downloadable, while the footnote says the authors will publish all source codes and data only after getting an acceptance letter. At the arXiv version I read, no working link is given. That matters because the reusable pipeline is the main practical value.\n\nSeparately, the paper's use of the literature is fair: the SBR-SVM baseline, the PD Net architectures, and the Ref. [22] thresholds are properly cited. I see no invented entities and no fitted parameters tuned to force the ranking. The reported statistics look honestly computed.\n\nWho is this for? Teams wanting a concrete example of how to evaluate interpretation methods on dopamine-transporter SPECT and how to use interpretation quality to choose between models. They will get a reusable procedure and a baseline to argue from. The specific winner, guided backpropagation, is plausible but not definitive, so a reader should treat the ranking as provisional.\n\nRecommendation: send to peer review. A serious referee should ask for (a) validation or sensitivity analysis of the striatal mask, (b) a quantitative measure for the SHAP claim, or else explicitly downgrade it to an observation, and (c) working code. These are fixable, and the paper deserves referee time.","headline":"A useful tutorial-style comparison of six XAI methods on SPECT-PD, with a plausible but not fully established ranking; the main soft spot is the unvalidated striatal ground-truth mask.","tokens_in":20010,"tokens_out":2113,"would_cite":false,"duration_ms":23165,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that among six interpretation methods for SPECT-based Parkinson's recognition, guided backpropagation best localizes the striatum (highest Dice coefficient) while SHAP best discriminates PD from normal at the uptake…","keywords":["Parkinson's disease","SPECT imaging","interpretable AI","deep convolutional neural networks","guided backpropagation","SHAP","model interpretation","computer-aided diagnosis"],"falsifier":"Replace the threshold-derived striatal masks with ground truth drawn independently by several nuclear-medicine physicians on the same scans and recompute the Dice rankings; if the ordering of interpretation methods changes, the paper's conclusion is an artifact of the threshold choice. A second check would repeat the comparison on an external SPECT dataset from a different scanner or preprocessing pipeline, since all images here come from a single acquisition protocol.","tokens_in":18925,"feed_emoji":"🧠","tokens_out":10140,"duration_ms":81778,"temperature":0.7,"pith_summary":"This tutorial paper answers a practical question: when a deep network classifies dopamine-transporter SPECT images as Parkinson's disease (PD) or normal control, which interpretation method should a clinician trust? The authors train four convolutional architectures on 607 subjects from a public SPECT dataset, apply six interpretation methods, and evaluate each method's attention map against a threshold-based striatum segmentation used as ground truth. They find that guided backpropagation yields the best fine-grained importance maps, shown by the highest Dice coefficient and lowest error, while SHAP best localizes the dopamine-uptake depletion in PD subjects. They also show that such interpretation scores can serve as a tie-breaker when choosing among models whose classification accuracies are statistically indistinguishable.","feed_headline":"Guided backpropagation gives clearest Parkinson's SPECT heatmaps","feed_subtitle":"In a six-method comparison, guided backprop best matches the striatum and SHAP best finds dopamine loss.","key_machinery":"The evaluative machinery is a thresholded segmentation of striatal nuclei that serves as ground truth for interpretation quality. Slices 35 through 48 of each SPECT volume are averaged and normalized, then segmented using mean uptake thresholds of 0.63 for normal controls and 0.69 for PD subjects, values taken from an earlier shape-analysis study; each interpretation method's attention map is thresholded at its top 10% and top 1% contribution pixels, and compared against this mask with the Dice coefficient, mean absolute error plots, and overlays of mean binary heatmaps. This turns the qualitative question 'does the network look at the right place' into a repeatable numeric ranking of interpretation methods, and the ranking is then used as feedback to select among otherwise tied classifiers.","core_discovery":"On the paper's own terms, the central discovery is that interpretation methods are not interchangeable for SPECT-based PD recognition: guided backpropagation dominates the other five tested methods in aligning attention with the striatal nuclei, reaching a mean Dice coefficient of 66.07 ± 12.62 at the top-1% threshold on Deep PD Net versus 45.32 for the next best method, and displaying the lowest mean absolute error inside the striatum. SHAP, though weaker on Dice, is the method whose mean binary heatmap visually marks the dopamine-depletion region in PD subjects, making it the best discriminator between PD and control at the clinically meaningful location. The authors further claim that these interpretation scores can break ties in model selection, and they recommend Deep PD Net for this task because its guided-backpropagation interpretation is significantly better than that of PD Net with batch normalization despite similar classification metrics.","pith_inferences":["A ground-truth mask weighted toward the dorsal putamen, the region affected earliest in PD, could shift the ranking toward SHAP rather than guided backpropagation, since SHAP's advantage lies exactly at the depletion site.","The conclusions come from two network families; re-running the evaluation on the more recent state-of-the-art architecture with thousands of images would test whether the ranking generalizes to higher-performing models.","The fixed thresholds from the cited shape-analysis study could be re-estimated per subject from striatal binding ratios, making the evaluation self-contained and potentially reducing segmentation bias."],"forward_implications":["Use guided backpropagation to verify at voxel level that a SPECT-PD classifier is focusing on the striatum rather than on artifacts.","Use SHAP to visualize where dopamine uptake is depleted in a given PD patient for clinical communication.","Interpretation scores can settle model selection when classifiers have statistically indistinguishable accuracy and specificity.","The evaluation procedure, Dice against a thresholded anatomical mask, transfers to other imaging CAD tasks with known anatomical targets."],"supporting_citations":[{"why":"Supplies the mean uptake thresholds (0.63 NC, 0.69 PD) used to segment the ground-truth striatal region.","marker":"[22]"},{"why":"Defines the PD Net architecture and training parameters from which all four tutorial networks are derived.","marker":"[38]"},{"why":"Introduces guided backpropagation, the interpretation method that scores highest on Dice in this evaluation.","marker":"[52]"},{"why":"Introduces SHAP, the interpretation method that best localizes the dopamine-depletion region in PD.","marker":"[55]"},{"why":"Defines DeepLIFT and proposes the top-percentage thresholding strategy that the paper adapts for evaluating attention maps.","marker":"[50]"},{"why":"Defines Grad-CAM, whose low-resolution last-layer maps the paper identifies as a cause of poor alignment.","marker":"[54]"},{"why":"Presents the state-of-the-art model the paper suggests for further study of the selection procedure.","marker":"[41]"},{"why":"Documents the public SPECT dataset from which all images in the tutorial are drawn.","marker":"[45]"}],"fun_headline_variants":["Guided backprop maps Parkinson's striatum best in SPECT","SHAP beats others at spotting dopamine loss in PD SPECT","Interpretation method choice matters for Parkinson's SPECT AI","Guided backprop and SHAP excel for Parkinson's SPECT interpretation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that striatal uptake regions as defined by fixed thresholds from a different study are the correct target that a faithful explanation should highlight; if a model attended to other clinically meaningful cues, or if those thresholds mis-segment these particular subjects, the Dice scores and the recommendation of guided backpropagation would not reflect true interpretation fidelity.","fun_headline_variants_meta":{"raw":{"variants":["Guided backprop maps Parkinson's striatum best in SPECT","SHAP beats others at spotting dopamine loss in PD SPECT","Interpretation method choice matters for Parkinson's SPECT AI","Guided backprop and SHAP excel for Parkinson's SPECT interpretation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000516,"raw_usage":{"total_tokens":2518,"prompt_tokens":975,"completion_tokens":1543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":1470}},"tokens_in":591,"tokens_out":1543,"duration_ms":11759,"temperature":1.0,"reasoning_tokens":1470,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:30:35.465860+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the threshold-derived striatal masks with ground truth drawn independently by several nuclear-medicine physicians on the same scans and recompute the Dice rankings; if the ordering of interpretation methods changes, the paper's conclusion is an artifact of the threshold choice. A second check would repeat the comparison on an external SPECT dataset from a different scanner or preprocessing pipeline, since all images here come from a single acquisition protocol.","supporting_citations":[{"cited_title":"High-accuracy classification of Parkinson’s disease through shape analysis and surface fitting in 123i-ioflupane SPECT imaging,","cited_arxiv_id":null,"evidence_quote":"Supplies the mean uptake thresholds (0.63 NC, 0.69 PD) used to segment the ground-truth striatal region."},{"cited_title":"Refining diagnosis of Parkinson’s disease with deep learning-based interpretation of dopamine transporter imaging,","cited_arxiv_id":null,"evidence_quote":"Defines the PD Net architecture and training parameters from which all four tutorial networks are derived."},{"cited_title":"A unified approach to interpreting model predictions,","cited_arxiv_id":null,"evidence_quote":"Introduces SHAP, the interpretation method that best localizes the dopamine-depletion region in PD."},{"cited_title":"Learning important fea- tures through propagating activation differences,","cited_arxiv_id":null,"evidence_quote":"Defines DeepLIFT and proposes the top-percentage thresholding strategy that the paper adapts for evaluating attention maps."},{"cited_title":"Grad-cam: Visual explanations from deep networks via gradient-based localization,","cited_arxiv_id":null,"evidence_quote":"Defines Grad-CAM, whose low-resolution last-layer maps the paper identifies as a cause of poor alignment."},{"cited_title":"An easy-to-use deep-learning model for highly accurate diagnosis of parkinson’s disease using spect images,","cited_arxiv_id":null,"evidence_quote":"Presents the state-of-the-art model the paper suggests for further study of the selection procedure."},{"cited_title":"The parkinson progression marker initiative (PPMI),","cited_arxiv_id":null,"evidence_quote":"Documents the public SPECT dataset from which all images in the tutorial are drawn."}],"review_version":1}