{"id":"e64d4c41-7c94-4d03-bd6e-4f5d117a3484","arxiv_id":"2502.09804","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A benchmark of four pretrained deep learning models for binary ALL classification reports 99.7% accuracy, but the paper's stage and hematogone detection claims are artifacts of label merging.","lead":"This paper tests four standard deep learning models, YOLOv8, YOLOv11, ResNet50, and Inception-ResNet-v2, for spotting acute lymphoblastic leukemia in microscope images of blood. The best model reports 99.7% accuracy on a merged public dataset, but the claimed ability to identify leukemia stages and hematogones is not supported by the binary design.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage and hematogone detection claims are impossible with a binary Normal/Cancer label design; the model output layer cannot produce stage or hematogone predictions.","rationale":"The reader's stated weakest assumption concerns HSV threshold sensitivity and data leakage, which are valid reproducibility and generalization concerns. However, the most load-bearing issue is the logical mismatch between the binary label design and the claimed multi-class capabilities. The reader's rationale does mention this mismatch as the primary reason for REJECT, so there is agreement on the verdict, but our stress test focuses on the internal inconsistency rather than the preprocessing details. The paper's reported binary accuracies are plausible and consistent with prior work, but the central contribution as stated in the abstract and contributions list is stage and hematogone identification, which cannot be delivered by a two-class model. No code or data are provided to check whether some other multi-class training occurred, so the claims are empirically unverifiable as written. A concrete, minimal test would be to inspect the output layer or the confusion matrices; if they are binary, the overclaim is confirmed. The verdict should remain REJECT because the main claimed novel capabilities are unsupported by the described experimental design.","tokens_in":12791,"tokens_out":2299,"duration_ms":22799,"concrete_test":"Inspect the output layer dimensions of the saved models or the code used for training. If the final classification layer has 2 output units, the model cannot distinguish ALL stages or hematogones as separate classes. Additionally, search the manuscript for any multi-class confusion matrix with rows for Early, Pre, Pro, Hematogone, and other Normal; if none exists, the stage and hematogone claims are unsupported and should be retracted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central contribution claims that the models 'can identify different ALL stages, including early stages' and 'can detect hematogones,' but the experimental design makes these claims impossible. The final merged dataset, as described in Table III, collapses Early, Pre, Pro, and Cancer from both sources into a single 'Cancer' class, and collapses Benign (including Hematogones) with ALL-IDB1 Normal into a single 'Normal' class. Training and evaluation are performed on these two classes only, and every reported confusion matrix (Figs. 8, 15, 22, 25, 28) is binary. A binary classifier cannot output a stage label (Early/Pre/Pro) or a distinct hematogone label, because those categories do not exist in the training targets or the output layer. Thus the stage and hematogone identification claims are not merely unverified; they are contradicted by the paper's own methodology. This is an internal inconsistency, not a generalization or robustness issue. The 99.7% accuracy may be credible as a binary cancer/normal benchmark, but the paper's novel contributions—stage detection and hematogone detection—are unsupported by the described architecture and labels.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a binary classification pipeline for Acute Lymphoblastic Leukemia (ALL) detection using YOLOv11n, YOLOv11s, YOLOv8s, ResNet50, and Inception-ResNet-v2. The authors merge two multi-cell datasets (ALL image dataset and ALL-IDB1), segment white blood cells with HSV thresholding, and train with transfer learning and data augmentation. They report test accuracies ranging from 97.3% (YOLOv11n) to 99.7% (Inception-ResNet-v2). The abstract and contributions additionally claim that the models can identify different ALL stages, including early stages, and can detect hematogones.","tokens_in":12984,"tokens_out":3413,"duration_ms":33007,"significance":"If the binary results are reproducible, the paper would provide a useful comparison of several deep-learning architectures on a multi-cell ALL detection task, a more realistic setting than single-cell benchmarks. The reported accuracies are plausible and the integration of two datasets is a sensible step toward robustness. However, the headline claims about stage identification and hematogone detection are not supported by the described experimental design, and the absence of segmentation threshold values and split details prevents replication. The contribution is therefore substantially weaker than the abstract suggests, though the underlying empirical comparison may still be of some value.","major_comments":[{"comment":"The central claims that the models \"can identify different ALL stages, including early stages\" and \"can detect hematogones\" are contradicted by the paper's own experimental design. Section II states that the merged dataset collapses Early, Pre, Pro, and Cancer into a single \"Cancer\" class, and Benign (including Hematogones) together with ALL-IDB1 Normal into a single \"Normal\" class. All reported confusion matrices (Figs. 8, 15, 22, 25, 28) are binary. A binary classifier has no output units for stage labels or for a separate hematogone class, so these claims are impossible given the described architecture and labels. The authors must either remove these claims or design a multi-class experiment with separate labels for each stage and for hematogones.","section":"§II (Table III), Abstract, Contributions"},{"comment":"The image segmentation step depends on hand-set HSV thresholds, but the threshold values are never reported. The text says only that \"upper and lower thresholds were set for the purple hue of the white blood cells.\" Without these values, the preprocessing is not reproducible, and the claim that the results generalize to \"real-world scenarios\" is unverifiable. The authors should provide the exact threshold values and discuss their sensitivity to staining, illumination, and dataset origin.","section":"§III-A"},{"comment":"The description of the train/validation/test split (70/15/15) is incomplete. The authors do not state whether the split was done at the image level, whether stratification was applied across the two merged datasets, or whether images from the same patient or source could appear in multiple splits. If image-level leakage occurred, the reported accuracies could be inflated. Specifics on the split procedure are needed to assess the validity of the performance numbers.","section":"§III-A"},{"comment":"All accuracy, F1, precision, recall, and specificity values are reported as point estimates from a single training run, with no error bars, confidence intervals, or multiple seeds. Given that ALL-IDB1 contributes only 108 images and the test set is a 15% subset of the merged data, the observed differences between models (e.g., 97.3% vs. 99.7%) may not be statistically significant. The paper should include repeated runs or at least confidence intervals to support comparative claims.","section":"§V–§VIII and Table IV"}],"minor_comments":[{"comment":"The sentence \"This mask was then applied to the original images ??, allowing the segmentation of the white blood cells\" contains a stray \"??\" placeholder and is grammatically incomplete; it should be corrected.","section":"§III-A"},{"comment":"In the comparison of YOLOv11s and YOLOv11n, the text states the improvement is \"0.9\" without specifying units; it should read \"0.9 percentage points\" to avoid ambiguity.","section":"§V"},{"comment":"The conclusion refers to \"YOLO11vs\" and \"YOLO8vs,\" which are inconsistent with the model names \"YOLOv11s\" and \"YOLOv8s\" used elsewhere in the paper.","section":"§XI"},{"comment":"Several figure captions are generic (e.g., \"Training and Validation Accuracy\") and do not clearly identify which model and configuration the figure corresponds to; this makes it harder to map the results to the text.","section":"§V–§VIII"},{"comment":"The comparison in Table V mixes results from different datasets (C-NMC 2019, ALL-IDB1, ALL dataset) without noting that direct accuracy comparison across datasets is not meaningful; a clarifying statement is needed.","section":"§X (Table V)"}],"recommendation":"reject","confidential_remarks":"The paper overstates its contributions: the stage and hematogone detection claims are not just unverified but logically incompatible with the binary label design. The missing segmentation thresholds and split details further undermine reproducibility. The binary accuracy results may be salvageable in a resubmission that reframes the paper as a multi-cell binary ALL classification benchmark and adds the missing methodological details, but as submitted the load-bearing claims cannot be fixed by minor edits."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the 99.7% binary accuracy is plausible and the dataset merge is a reasonable idea, but the abstract's claims about identifying ALL stages and detecting hematogones are not supported by the experimental design. The final label space, shown in Table III, collapses Early/Pre/Pro/Cancer into Cancer and Benign/Normal into Normal. A binary classifier cannot output stage or hematogone labels. That is an internal inconsistency, not a generalization worry.\n\nWhat's genuinely useful: they train on multi-cell images from two public datasets, which is closer to real smears than the single-cell datasets common in this literature. The accuracy numbers across five models are internally consistent, and Inception-ResNet-v2 at 99.7% with 100% specificity on the test set is a concrete, checkable result. The comparison to prior work in Table V is a useful summary.\n\nSoft spots, in rough order: (1) the stage/hematogone framing is the main contribution and it cannot be delivered by the architecture; the paper would need a multi-class model or a separate hematogone class to support those sentences. (2) No error bars, no confidence intervals, and the test set is a 15% split of a merged dataset, so the 99.7% is a point estimate on a few hundred images. (3) The HSV segmentation thresholds are never reported, which makes the preprocessing unreproducible. (4) The YOLOv8 description is factually wrong (EfficientNet backbone, NAS-FPN, Focal Loss, Mixup, APAS are not YOLOv8 as released by Ultralytics). (5) Table V compares across different datasets as if the numbers were commensurate; it should say so explicitly.\n\nI don't think the authors are hiding anything—Table III is transparent about the merging. The problem is that the abstract and conclusion go beyond what the labels can support. That is a fixable framing error, not a fraudulent result. As submitted, I would not cite it, and I'd be hesitant to put it in front of students as a model of careful claim-making. But the underlying binary benchmark could be a legitimate contribution if reframed and made reproducible.\n\nRecommendation: send to peer review only if the editor expects a major revision that removes the stage/hematogone claims, reports error bars, and releases code and thresholds. Otherwise desk reject as an incremental benchmark with overreach.","headline":"A plausible binary ALL/normal benchmark whose stage and hematogone claims are impossible by construction; worth a referee only to force an honest reframing.","tokens_in":13565,"tokens_out":2170,"would_cite":false,"duration_ms":21923,"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":"A fine-tuned Inception-ResNet-v2 distinguishes acute lymphoblastic leukemia from normal blood cells with 99.7% accuracy.","keywords":["Acute Lymphoblastic Leukemia","YOLOv11","YOLOv8","ResNet50","Inception-ResNet-v2","image segmentation","hematogones","transfer learning"],"falsifier":"Reproduce the segmentation using the exact HSV thresholds, which the paper does not report, on blood-smear images from a third laboratory; if the binary masks lose parts of blast nuclei or retain platelet fragments, then the near-perfect test scores are bound to the two source datasets rather than to a generalizable rule.","tokens_in":12537,"feed_emoji":"🩸","tokens_out":12173,"duration_ms":100563,"temperature":0.7,"pith_summary":"This paper claims that deep learning models fine-tuned on multi-cell blood-smear images can tell malignant white blood cells from healthy ones reliably enough for real-world leukemia screening. The authors merge two public multi-cell datasets, the ALL image dataset and ALL-IDB1, into Normal and Cancer classes, so the models cannot rely on single-cell crops. The reported results put Inception-ResNet-v2 at 99.7% accuracy with 100% specificity, with YOLOv11, YOLOv8, and ResNet50 in the 97–99% range. A sympathetic reader would care because the method is aimed at early stages of ALL and at hematogones, the benign lookalike cells that often cause false ALL diagnoses.","feed_headline":"Fine-tuned model spots leukemia in blood smears at 99.7%","feed_subtitle":"On multi-cell blood smears, Inception-ResNet-v2 hit 99.7% accuracy with zero normal cells mislabeled as cancer.","key_machinery":"The argument is carried by a two-stage pipeline. In the first stage, blood-smear images are converted to HSV color space (hue, saturation, value), and a threshold on the purple hue of white blood cells builds a binary mask that removes backgrounds, platelets, and red cells; the exact threshold values are not reported. In the second stage, pretrained YOLOv8, YOLOv11, ResNet50, and Inception-ResNet-v2 are fine-tuned on the masked images, with the two ResNet-family classifiers having their last ten layers unfrozen and using SGD at a 0.001 learning rate with augmentation such as flipping, rotation, zoom, and mosaic. This transfer-learning step is what lets a modest merged dataset reach the reported accuracies above 97%.","core_discovery":"On the paper's own terms, the discovery is that fine-tuning large convolutional models on multi-cell blood-smear images yields near-perfect binary ALL classification. Inception-ResNet-v2 records 99.7% test accuracy with 100% precision and 100% specificity, and ResNet50 records 99% accuracy with 100% recall; the YOLO detectors trail slightly, with YOLOv11s at 98.2% accuracy and YOLOv11n at 97.3%. Because the Normal class contains benign cells and hematogones while the Cancer class contains early, pre, and pro stages, the authors read these metrics as evidence that the models recognize early-stage disease and resist the standard hematogone false-positive trap.","pith_inferences":["A natural stress test is to run the same unreported HSV thresholds on blood-smear images from a third laboratory; the outcome would show whether the 99.7% figure transfers across staining and illumination conditions.","The per-model trade-offs are uneven, so a deployed system would likely pair a high-specificity classifier like Inception-ResNet-v2 with a high-recall classifier like ResNet50 rather than rely on one model.","A further testable extension is to convert the multi-cell binary decision into cell-level counts, using the number of blast cells a model misses per slide to set a clinical alert threshold."],"forward_implications":["A blood-smear slide can be classified as malignant or benign in a single pass, because the models were trained and tested on multi-cell images rather than isolated cells.","Early, pre, and pro stages are folded into the Cancer class, so a correct positive prediction can in principle catch ALL before blast cells dominate a sample.","Hematogones are treated as Normal, and the 100% specificity of Inception-ResNet-v2 on the test set means these lookalike cells produce almost no false alarms.","Because the merged dataset combines images from two sources, the reported accuracies support the authors' claim that the method is not tied to a single imaging setup."],"supporting_citations":[{"why":"It supplies the ALL image dataset's 3,256 multi-cell images, including the benign, early, pre, and pro classes that define the Cancer and Normal labels.","marker":"[6]"},{"why":"It supplies the ALL-IDB1 microscopy images whose normal and cancer samples are merged with the ALL image dataset.","marker":"[12]"},{"why":"It introduces YOLOv11, the detection architecture the paper applies for the first time to blood cancer.","marker":"[8]"},{"why":"It provides the YOLOv8 model used as the comparison detector.","marker":"[9]"},{"why":"It provides the ResNet50 architecture with residual connections that the paper fine-tunes.","marker":"[10]"},{"why":"It provides the Inception-ResNet-v2 architecture that records the highest accuracy.","marker":"[18]"},{"why":"It documents the authors' earlier YOLOv8 and YOLOv11 results that this study extends to four models on the merged dataset.","marker":"[11]"},{"why":"It supplies the transfer-learning and fine-tuning definitions that motivate the training procedure.","marker":"[16]"}],"fun_headline_variants":["AI model hits 99.7% accuracy detecting leukemia in blood smears","Deep learning spots early-stage leukemia at 99.7% accuracy","YOLO and ResNet models diagnose leukemia with 99.7% accuracy","Leukemia detection: Inception-ResNet-v2 reaches 99.7% accuracy","New AI detects leukemia and benign mimics at 99.7% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result rests on a fixed set of hand-picked color thresholds that separate white blood cells from everything else in every image; if those thresholds are sensitive to staining, lighting, or dataset origin, the near-perfect accuracy will not hold on new slides.","fun_headline_variants_meta":{"raw":{"variants":["AI model hits 99.7% accuracy detecting leukemia in blood smears","Deep learning spots early-stage leukemia at 99.7% accuracy","YOLO and ResNet models diagnose leukemia with 99.7% accuracy","Leukemia detection: Inception-ResNet-v2 reaches 99.7% accuracy","New AI detects leukemia and benign mimics at 99.7% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000687,"raw_usage":{"total_tokens":3098,"prompt_tokens":916,"completion_tokens":2182,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":2080}},"tokens_in":532,"tokens_out":2182,"duration_ms":13136,"temperature":1.0,"reasoning_tokens":2080,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T20:25:48.702978+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the segmentation using the exact HSV thresholds, which the paper does not report, on blood-smear images from a third laboratory; if the binary masks lose parts of blast nuclei or retain platelet fragments, then the near-perfect test scores are bound to the two source datasets rather than to a generalizable rule.","supporting_citations":[{"cited_title":"Inception-v4, inception-resnet and the impact of residual connections on learning,","cited_arxiv_id":null,"evidence_quote":"It provides the Inception-ResNet-v2 architecture that records the highest accuracy."},{"cited_title":"Early diagnosis of acute lymphoblastic leukemia using yolov8 and yolov11 deep learning models,","cited_arxiv_id":null,"evidence_quote":"It documents the authors' earlier YOLOv8 and YOLOv11 results that this study extends to four models on the merged dataset."},{"cited_title":"A fast and efficient cnn model for b-all diagnosis and its subtypes classification using peripheral blood smear images,","cited_arxiv_id":null,"evidence_quote":"It supplies the ALL image dataset's 3,256 multi-cell images, including the benign, early, pre, and pro classes that define the Cancer and Normal labels."},{"cited_title":"DL4ALL: Multi-Task Cross-Dataset Transfer Learning for Acute Lymphoblastic Leukemia Detection,","cited_arxiv_id":null,"evidence_quote":"It supplies the ALL-IDB1 microscopy images whose normal and cancer samples are merged with the ALL image dataset."},{"cited_title":"Yolov11 - key features,","cited_arxiv_id":null,"evidence_quote":"It introduces YOLOv11, the detection architecture the paper applies for the first time to blood cancer."},{"cited_title":"Yolov8: A novel object detection algorithm with enhanced performance and robustness,","cited_arxiv_id":null,"evidence_quote":"It provides the YOLOv8 model used as the comparison detector."},{"cited_title":"A review of deep transfer learning and recent advancements,","cited_arxiv_id":null,"evidence_quote":"It supplies the transfer-learning and fine-tuning definitions that motivate the training procedure."}],"review_version":1}