{"id":"da7bf431-f338-4443-9108-c581389f119b","arxiv_id":"2506.07652","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FMaMIL combines Mamba-based multiple instance learning with learnable frequency-domain encoding and CAM-guided pseudo-label refinement to segment lesions from image-level labels only.","lead":"This paper proposes FMaMIL, a two-stage weakly supervised system that segments lesions in pathology images using only image-level labels, combining Mamba-based sequence modeling with learnable frequency-domain features. The authors report state-of-the-art results on two histopathology datasets, including CAMELYON16, without using pixel-level annotations during training.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CAMELYON16 preprocessing may assign per-patch labels from slide-level labels, converting the claimed MIL training into noisy patch classification and potentially invalidating the headline comparison.","rationale":"I agree with the reader's weakest assumption; this is the load-bearing issue because it determines whether the experiment tests the proposed method. If patch labels are used, the method reduces to a supervised patch classifier, so the claim of a Mamba-based MIL framework is unverified; if patches from the same WSI leak across splits, segmentation metrics are optimistic. I do not change the reader's conditional verdict because the ambiguity can be resolved by inspecting the code and rerunning the protocol. Secondary concerns (validation-set reporting, hyperparameter tuning) also matter but are less decisive.","tokens_in":27577,"tokens_out":7145,"duration_ms":87195,"concrete_test":"Examine the released code at https://github.com/chenghangbei0702/FmaMIL and the CAMELYON16 data loader: (1) determine whether a binary patch label is created from the slide label and used in any loss term; (2) check whether training and validation sets are split at WSI level or patch level. Then rerun the first stage using only bag-level labels (one label per WSI) with a strict WSI-exclusive split, keeping all other hyperparameters fixed, and compare the resulting mIoU/Dice on CAMELYON16 against Table 2. If the gap over MambaMIL substantially shrinks or vanishes, the current numbers are not evidence for the proposed MIL architecture.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim rests on the CAMELYON16 experiment, but the preprocessing described in §4.1 is inconsistent with the MIL formulation used in §3.1.2 and §3.5. The text states: 'The label for each patch was assigned based on the patient-level label, ensuring label consistency and accuracy.' Taken literally, every patch from a positive slide is labeled positive and every patch from a negative slide is labeled negative. This is a noisy fully-supervised patch classification setup, not MIL: instance labels are no longer unknown, and the bag-level attention aggregation (Eqs. 17–20) is no longer the only source of supervision. If those per-patch labels are fed into any loss (e.g., a patch-level cross-entropy), then FMaMIL has access to far more label signal than the MIL baselines (MambaMIL, AB-MIL, Trans-MIL) which use only one label per bag. The reported 0.869/0.957 Dice on CAMELYON16 could then be explained by this supervision shortcut rather than by the Mamba or frequency modules. The paper never specifies whether patch labels are used in training, in the instance-level loss (Eq. 23), or only for constructing bags. Additionally, Table 1's train/test/val patch counts suggest a patch-level split; if patches from the same WSI appear in both training and validation, segmentation metrics are inflated by overlap. This makes the headline comparison to SOTA weakly supervised methods uninterpretable without clarification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FMaMIL, a two-stage weakly supervised framework for lesion segmentation in histopathology. Stage one extracts patch features with a pre-trained CNN, applies a Mamba-based encoder with bidirectional spatial scanning and a learnable frequency-domain encoding module, and trains an attention-based MIL head with both bag-level and instance-level losses. Stage two generates CAMs from this head, binarizes them into pseudo-labels, and trains a U-Net with a combination of soft and hard losses, plus a confidence-based pseudo-label refinement and consistency loss. The method is evaluated on a private glomerular-lesion dataset and on CAMELYON16, with claims of state-of-the-art weakly supervised segmentation performance and, surprisingly, performance above fully supervised U-Net and DeepLabV3+.","tokens_in":27859,"tokens_out":7456,"duration_ms":90382,"significance":"If the reported results are valid, the architectural combination—Mamba-based long-range modeling, learnable frequency encoding, and a two-stage pseudo-label refinement scheme—would be a useful contribution to weakly supervised histopathology segmentation, and the public code link would aid reproducibility. The core ideas are plausible and the problem is practically important. However, the current evidence is not sufficient to support the head-line claims: the CAMELYON16 preprocessing described in Section 4.1 conflicts with the MIL assumptions of Section 3.1.2, the instance-level loss in Eq. (23) is not well-defined, and the evaluation protocol tunes hyperparameters on the same validation set used for final comparisons. These issues are central to the empirical claims rather than superficial presentation defects.","major_comments":[{"comment":"The CAMELYON16 description states that \"the label for each patch was assigned based on the patient-level label,\" which directly contradicts the MIL formulation of Section 3.1.2 (Eq. (5)) where instance labels are latent and only bag labels are observed. If these per-patch labels are used in any training loss, such as the instance-level loss in Eq. (23), FMaMIL receives noisy dense per-patch supervision rather than bag-level supervision. That would make the comparison with bag-level baselines (AB-MIL, Trans-MIL, MambaMIL) invalid as a test of the proposed MIL or frequency modules. The paper must clarify exactly how patch labels are used in each stage, and if they are used only to construct bags, state this explicitly.","section":"§4.1, §3.1.2"},{"comment":"Equation (23) defines L_instance = -Σ_i α_i log ŷ_i but contains no ground-truth or soft target for the instance. As written, minimizing this term drives every instance prediction toward 1, since α_i is nonnegative and log ŷ_i is maximized when ŷ_i = 1. If ŷ_i is a probability vector, the expression is vector-valued rather than a scalar loss. The instance-level supervision term must be rewritten with an explicit target and a consistent scalar formulation.","section":"§3.5, Eq. (23)"},{"comment":"The reported numbers are obtained on the validation set (Section 4.2: \"the best accuracy on the validation set was reported\"), and the same validation set is used to select the patch size (Table 3), the loss balance λ (Fig. 11), the threshold t (Fig. 12), and the loss balance α (Fig. 13). Tables 2 through 8 report single runs without error bars or significance tests. The reported performance gaps over baselines may therefore partly reflect selection on the evaluation set. The authors should report results on a separately held-out test split, or use cross-validation with hyperparameters fixed independently of the test set, and provide error bars or significance measures.","section":"§4.2, §4.4, Tables 2–8"},{"comment":"The patch statistics in Table 1 suggest that the train/test/validation split is performed at the patch level rather than at the whole-slide level. In CAMELYON16, for example, 13,184 training patches, 3,776 test patches, and 1,883 validation patches are derived from only 270 WSIs; if patches from the same WSI appear in both training and validation, the evaluation is inflated by spatial correlation. The paper should clarify whether the split is by WSI or by patch, and if the latter is the case, the experiments should be rerun with a WSI-exclusive split.","section":"§4.1, Table 1"},{"comment":"The pseudo-label refinement rule y_refined = β·y_cam + (1−β)·y_pred conflicts with the surrounding text. For β > 0.7, the CAM is weighted more strongly than the model prediction, yet the text states that \"in regions where the model's predictions have high confidence, the model's predictions are prioritized over the CAM labels.\" The parameter β is also described as a confidence threshold but used as a mixing weight. This inconsistency must be resolved to make the self-correction mechanism reproducible.","section":"§3.6, Eq. (27)"}],"minor_comments":[{"comment":"The introduction states that \"Section 4 presents experimental results on three datasets,\" but Section 4.1 describes only two datasets (private glomerular lesions and CAMELYON16). Please correct the count.","section":"§1"},{"comment":"The notation F''[-1] in Eq. (13) is not defined; it presumably denotes the reverse of the sequence, but this should be stated explicitly.","section":"§3.4.3, Eq. (13)"},{"comment":"Algorithm 1 line 5 (\"LearnableParameters\") duplicates the learnable-weight operation already introduced in line 3 and does not match the text around Eq. (11). Please align the algorithm box with the equations.","section":"§3.4.2, Algorithm 1"},{"comment":"The rows of Table 3 are difficult to read: the columns \"Patch size\" and \"Num_patchs\" contain combined values such as \"1100 22*22\" and the final row \"FMaMIL(100)\" is not aligned with the other rows. Please reformat the table with clear, separate columns.","section":"Table 3"},{"comment":"The figure captions contain typos: \"FSSS\" should be \"FSS\" or \"U-Net,\" and \"PatchCamelyon16\" should be \"CAMELYON16\" to match the dataset name used elsewhere in the text.","section":"Figures 14 and 15"},{"comment":"The use of absolute value in Eq. (14) is ambiguous because x_fft_space may be complex after the IFFT; please clarify whether the real part or the complex modulus is used.","section":"§3.4.3, Eq. (14)"}],"recommendation":"major_revision","confidential_remarks":"The reader's central concern about CAMELYON16 preprocessing is well-founded and should be resolved before publication. If per-patch labels are not used as supervision, the authors must state so explicitly and describe how the bags are constructed; if they are used, the comparison with weakly supervised baselines is not a fair test of MIL. The evaluation-protocol issues (validation-based hyperparameter selection and single-run results) also need to be addressed. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: FMaMIL is a coherent combination of existing ideas—Mamba MIL with row/column scans, a learnable FFT filter, and CAM-guided pseudo-label refinement. The parts are known, but putting them together for weakly supervised lesion segmentation is new, and the ablations suggest each piece helps. The bigger problem is not the architecture; it's the experimental protocol. Section 4.1 says on CAMELYON16 \"the label for each patch was assigned based on the patient-level label.\" That sentence could mean they simply tagged patches for bag construction, but it reads like they created per-patch labels. If those labels were used as any form of supervision, the method is no longer MIL, and the comparison to AB-MIL, Trans-MIL, MambaMIL etc. is unfair. The paper never specifies how those patch labels are used. Combined with Table 1, which lists positive/negative patch counts per split, the setup looks like patch-level classification rather than bag-level MIL. There is also the usual validation-set tuning problem: lambda, threshold, and alpha are all selected on the same validation set used to report final numbers, with single runs and no error bars. Patch-level splitting of WSIs could leak same-slide patches across splits and inflate Dice. So the headline numbers cannot be interpreted as a fair SOTA comparison until these are resolved.\n\nWhat is good: the method is well motivated, the figures are clear, and the ablations are systematic enough to indicate the frequency encoding contributes something beyond the Mamba baseline. They also include a private dataset, which tests generalization, and they provide a code link. The explanation for beating fully supervised methods—noisy pixel annotations—is plausible but needs independent verification.\n\nMy advice: this is a conditional. The architectural contribution is real but incremental; the evaluation could be sound if the authors clarify that patch labels are only used for bag assignment and then re-run with slide-level splits and a held-out test set. As is, I would not trust the SOTA comparison. Still, a serious referee can fix this; desk-rejecting would be a mistake. Send to peer review and ask for clarification and re-evaluation.","headline":"Useful architecture, untrustworthy CAMELYON16 protocol—needs clarification of patch-label usage and a clean held-out evaluation before the numbers mean anything.","tokens_in":28430,"tokens_out":2216,"would_cite":false,"duration_ms":26979,"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":"FMaMIL reports weakly supervised lesion segmentation from image-level labels alone, with mIoU/Dice of 0.887/0.934 on glomerular lesions and 0.869/0.957 on CAMELYON16.","keywords":["Frequency-domain","histopathology images","multiple instance learning","lesion segmentation","Mamba model"],"falsifier":"Train the identical pipeline on CAMELYON16 with true bag-level supervision, meaning no per-patch labels are assigned during preprocessing, and compare segmentation mIoU/Dice; if performance falls to the level of the ablated model without the frequency module, the reported gains came from the patch-label shortcut described in Section 4.1 rather than from the Mamba or frequency architecture.","tokens_in":27322,"feed_emoji":"🔬","tokens_out":10430,"duration_ms":120603,"temperature":0.7,"pith_summary":"This paper aims to show that lesion segmentation in histopathology can be driven by image-level labels alone, without pixel-level annotations. The proposed FMaMIL pipeline first uses a Mamba-based multi-instance encoder, boosted by learnable Fourier-domain filtering and bidirectional row/column scanning, to produce class activation maps (CAMs); then a U-Net is trained with CAM-based soft supervision, binarized hard labels, and a confidence-weighted self-correction loss. On a private glomerular-lesion dataset and the public CAMELYON16 dataset, the paper reports mIoU/Dice of 0.887/0.934 and 0.869/0.957 respectively, above the weakly supervised and fully supervised baselines in Table 2. The intended payoff is that diagnostic histopathology segmentation could be trained where only case-level diagnoses are available.","feed_headline":"Frequency-Mamba model tops lesion segmentation with image-level labels","feed_subtitle":"Reports mIoU/Dice of 0.887/0.934 and 0.869/0.957 on two histology datasets using only image-level labels.","key_machinery":"The load-bearing component is the FMamba block, a Mamba encoder layer with three parallel branches: row-scanned and column-scanned sequences, each processed by forward and backward selective state-space models and gated with a SiLU-activated branch, plus a learnable frequency-domain encoding that runs FFT along the sequence, applies learned complex frequency weights, runs IFFT, and adds a skip connection. The three branch outputs are summed and passed through a residual connection. This mechanism converts a sequence of image patches into a fused feature sequence in linear time, capturing long-range dependencies while injecting spectral texture and boundary information; the attention weights from the MIL classification head are then converted into class activation maps that seed second-stage segmentation training.","core_discovery":"The central claim is that combining a selective state-space model with frequency-domain information makes a weak multi-instance learner produce substantially better class activation maps, and that those maps, when refined by soft supervision and self-correction, are accurate enough to supervise a segmentation model. The paper reports classification accuracy/AUC of 0.996/0.998 on the glomerular dataset and 0.993/0.992 on CAMELYON16, with segmentation mIoU/Dice of 0.887/0.934 and 0.869/0.957 respectively, and states that these numbers exceed all weakly supervised competitors and also the fully supervised U-Net and DeepLabV3+ in Table 2. The gain is attributed to three cooperating modules: a learnable frequency-domain encoding that applies FFT, multiplies the spectrum by learned complex weights, and returns via IFFT with a skip connection; a bidirectional scanning strategy that models inter-patch context in both row and column orders; and an attention-based instance-to-bag classification head whose instance-level loss sharpens the attention maps that become CAMs.","pith_inferences":["Editorial inference: The CAMELYON16 preprocessing described in Section 4.1, where the label for each patch was assigned from the patient-level label, conflicts with the multi-instance premise that instance labels are unknown; if taken literally, the first stage could be learning from noisy per-patch labels inherited from positive slides, and the reported gains may not isolate the Mamba or frequenc","Editorial inference: The optimal patch size of 100x100 on 1100x1100 images is close to the scale of glomerular structures, so a natural testable extension is whether multi-scale or lesion-scale-normalized patch sampling preserves the gains on datasets with variable lesion sizes.","Editorial inference: The reported edge over fully supervised baselines may depend on incomplete pixel annotations in the evaluation sets; on a densely annotated dataset the fully supervised models might narrow the gap, so an evaluation with exhaustive annotations would sharpen the comparison."],"forward_implications":["Segmentation models for histopathology could be trained from case-level diagnoses alone in the evaluated settings, removing the need for pixel-level pathologist annotation.","On the paper's two test sets, the pipeline reports segmentation quality above fully supervised U-Net and DeepLabV3+, suggesting that weak supervision is not necessarily a ceiling relative to pixel-trained models when annotations are sparse or noisy.","Because the Mamba encoder is linear in sequence length, the approach scales to the long patch sequences of whole-slide images more efficiently than transformer-based multi-instance baselines.","The ablation attributes the largest single-stage gain to the learnable frequency-domain encoding, indicating spectral features carry much of the segmentation improvement independently of the scanning strategy."],"supporting_citations":[{"why":"Defines the baseline Mamba encoder and the training recipe (augmentation, ResNet50 features, optimizer, scheduling, EMA) that FMaMIL adopts.","marker":"[75]"},{"why":"Supplies the attention-based instance-to-bag pooling that produces the attention weights reused as soft instance-level supervision.","marker":"[24]"},{"why":"Is the Mamba-MIL baseline with sequence reordering that FMaMIL extends and compares against.","marker":"[67]"},{"why":"Introduces correlated MIL with self-attention, the main transformer-based baseline for inter-instance modeling.","marker":"[49]"},{"why":"Provides the CAMELYON16 dataset and its slide-level labels used in the evaluation.","marker":"[13]"},{"why":"Defines class activation maps, the pseudo-label source converted into soft and hard segmentation supervision.","marker":"[73]"},{"why":"Provides the selective state-space model equations that the FMamba block discretizes and uses for long-sequence modeling.","marker":"[17]"},{"why":"Supplies the pretrained ResNet50 used to extract per-patch features before the FMamba encoder.","marker":"[21]"},{"why":"Supplies the U-Net segmentation backbone of the second stage and the fully supervised baseline it is compared with.","marker":"[47]"},{"why":"Earlier frequency-plus-spatial medical segmentation model whose frequency enhancement and gating motivate the learnable frequency branch.","marker":"[20]"}],"fun_headline_variants":["Mamba plus frequency encoding lifts weak lesion segmentation to top","FMaMIL: Frequency-aware Mamba MIL outperforms fully supervised","Image-level labels only: FMaMIL sets lesion segmentation records","Frequency-Mamba MIL: weak labels, strong segmentation results","FMaMIL: Using FFT and Mamba to segment lesions without pixel labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's claim depends on the first stage learning from slide-level labels alone, but the paper says each patch in CAMELYON16 was assigned the label of its slide, so positive slides contribute many positively labeled patches; if that is what was done, the model may be doing ordinary supervised patch classification instead of multi-instance learning.","fun_headline_variants_meta":{"raw":{"variants":["Mamba plus frequency encoding lifts weak lesion segmentation to top","FMaMIL: Frequency-aware Mamba MIL outperforms fully supervised","Image-level labels only: FMaMIL sets lesion segmentation records","Frequency-Mamba MIL: weak labels, strong segmentation results","FMaMIL: Using FFT and Mamba to segment lesions without pixel labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000388,"raw_usage":{"total_tokens":2055,"prompt_tokens":961,"completion_tokens":1094,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":1002}},"tokens_in":577,"tokens_out":1094,"duration_ms":11887,"temperature":1.0,"reasoning_tokens":1002,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:28:56.589377+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical pipeline on CAMELYON16 with true bag-level supervision, meaning no per-patch labels are assigned during preprocessing, and compare segmentation mIoU/Dice; if performance falls to the level of the ablated model without the frequency module, the reported gains came from the patch-label shortcut described in Section 4.1 rather than from the Mamba or frequency architecture.","supporting_citations":[{"cited_title":"Weakly supervisedhistopathology cancerimage segmentationand classifica- tion","cited_arxiv_id":null,"evidence_quote":"Defines the baseline Mamba encoder and the training recipe (augmentation, ResNet50 features, optimizer, scheduling, EMA) that FMaMIL adopts."},{"cited_title":"Attention-based deep multiple instance learning, in: International conference on machine learning, PMLR","cited_arxiv_id":null,"evidence_quote":"Supplies the attention-based instance-to-bag pooling that produces the attention weights reused as soft instance-level supervision."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the Mamba-MIL baseline with sequence reordering that FMaMIL extends and compares against."},{"cited_title":"Transmil:Transformerbasedcorrelatedmultipleinstancelearningfor whole slide image classification","cited_arxiv_id":null,"evidence_quote":"Introduces correlated MIL with self-attention, the main transformer-based baseline for inter-instance modeling."},{"cited_title":"Learning deep features for discriminative localization, in: Proceed- ings of the IEEE conference on computer vision and pattern recogni- tion, pp","cited_arxiv_id":null,"evidence_quote":"Defines class activation maps, the pseudo-label source converted into soft and hard segmentation supervision."}],"review_version":1}