List of Papers Browse by Subject Areas Author List
Abstract
Breast cancer analysis requires integrating morphological assessment of the tumor microenvironment with molecular receptor status (ER, PR, HER2). Yet computational pathology datasets address these tasks in isolation, providing either nucleus annotations without clinical labels or molecular profiles without spatial cellular context. We present BC-MultiSet, the first histopathology dataset that bridges this gap by uniquely pairing pixel-level instance segmentation masks for three cell types (Tumor, Inflammatory, Fibroblasts) with corresponding slide-level clinical biomarkers (ER, PR, HER2). Derived from 50 invasive breast carcinoma patients, the dataset contains 5,000+ H\&E patches with 109,180 nuclei manually annotated by two board-certified pathologists via consensus. We establish six classification architectures (including Virchow2) and three instance segmentation models (CellViT++, Cerberus, HoVer-NeXt). We also introduce multi-task learning baselines that jointly model morphology and clinical targets, revealing both the promise and challenges of integrated analysis. BC-MultiSet enables, for the first time, direct computational investigation of how cellular morphology relates to molecular expression, providing a foundational resource for developing holistic digital pathology workflows. The code is available at https://github.com/LanaLana/BC-MultiSet. The dataset is available at https://doi.org/10.5281/zenodo.20717567.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4838_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/LanaLana/BC-MultiSet
Link to the Dataset(s)
BC-MultiSet dataset: http://doi.org/10.5281/zenodo.20717567
BibTex
@InProceedings{IllSve_BCMultiSet_MICCAI2026,
author = { Illarionova, Svetlana AND Studenikina, Anastasiia AND Mohnenko, Sergey AND Akaeva, Aida AND Akbaba, Elif AND Boyko, Maria AND Ekenel, Hazım Kemal AND Grinevich, Vyacheslav AND Hamoudi, Rifat AND Sharaev, Maxim},
title = { { BC-MultiSet: A Multi-target Dataset and Benchmark for Clinical Tasks in Breast Cancer } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper describes an annotated breast cancer dataset to train and evaluate models for segmentation and classification tasks. It demonstrates the use of the dataset with a number of deep learning architectures in segmentation and classification tasks.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- A good number of annotated nuclei from large number of tissue patches.
- The dataset has slide-level classifications (ER, PR, HER2) in addition to nucleus segmentation and labels.
- The use of the dataset is shown using several classification and segmentation models.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The number of patients from whom the tissue patches have been extracted is small. This could limit the use of the dataset in patient-level classification tasks.
- The experimental evaluation is done using the proposed dataset only. Evaluation with external test datasets (e. g. , TCGA) should be done to assess the strengths and limitations of the dataset.
- Some existing datasets such as PanNuke and NuCLS provide similar nucleus level annotations. An experimental comparison with those datasets would help better assess the contributions and limitations of the proposed dataset.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
A dataset with a good number of pixel level annotations and linked slide-level classification information. However, the experimental evaluation is limited to the proposed dataset only.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have addressed some of the critique. BC-MultiSet can be useful for development and evaluation of image analysis methods, despite the limited comparison with other datasets in the paper.
Review #2
- Please describe the contribution of the paper
The paper’s primary contribution is a per-nucleus annotated dataset linking instance segmentation masks covering tumor, inflammatory, and fibroblast cell populations to molecular clinical labels (ER, PR, HER2) across 50 WSIs from 50 patients. To contextualize the dataset’s utility, the authors provide a benchmark of classification and instance segmentation models trained and evaluated on this data. They further propose joint multi-task baselines that couple segmentation and classification into a single learning pipeline, offering a reference point for future work on this dataset.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper is very straightforward. The central contribution of this work is the dataset, with the remainder of the paper dedicated to experimental insights that characterize its scope and potential utility. The fine-grained nucleus-level annotations, validated by board-certified pathologists and coupled with molecular clinical labels, address a well-recognized bottleneck in computational pathology — the scarcity of high-quality, clinically grounded annotation resources.
The authors compare their dataset to well-known ones like MoNuSeg, NuCLS, and PanNuke to show how it fits in. They claim their dataset is the only one with both nucleus annotations and clinical labels for the same patients. The proposed dataset reports the baseline performance of state-of-the-art classifiers and segmentation models, establishing an initial benchmark for the community on their dataset.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper does not present any methodological innovation. The experimental framework relies entirely on existing classifier and segmentation architectures, with no novel algorithmic contributions introduced. The work is positioned purely as a dataset contribution, and the accompanying experiments serve as benchmarks rather than methodological advancements.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
1.The authors should clarify their motivation for focusing specifically on breast cancer and breast tissue annotations. Furthermore, a discussion comparing breast tissue annotation with other organ types would strengthen the paper. What kind of challenges and variability other organs might impose compared to breasts?
2.How much time did it consume per slide for manual annotation and verification by pathologists?
3.The caption accompanying Figure 2 lacks a color legend, making it difficult to identify which color corresponds to which cell category (tumor, inflammatory, or fibroblast). Furthermore, the figures are not referenced within the manuscript text.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Despite the absence of methodological novelty, the paper makes a meaningful contribution to the image analysis and computer vision communities through the introduction of a well-curated, clinically grounded dataset. The value of such a resource should not be understated. Fine-grained, nucleus-level annotations paired with molecular clinical labels are scarce, and this dataset has strong potential to facilitate significant future method development in computational pathology.
A particularly notable strength is the annotation verification process, in which two board-certified pathologists validated the nucleus-level labels. This lends a high degree of clinical credibility and reliability to the dataset.
The experimental section, while straightforward, provides genuinely useful insights into dataset characteristics and model behavior.
Finally, the public release of both the dataset and the associated codebase is commendable, as it ensures full reproducibility and lowers the barrier for adoption by the broader research community.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I believe the authors addressed most of the concerns raised.
Review #3
- Please describe the contribution of the paper
The paper introduces BC-MultiSet, a breast cancer H&E histopathology dataset of 5,586 patches from 50 invasive-carcinoma patients, with 109,180 manually annotated nuclei across three classes (tumor, fibroblast, inflammatory) resolved by consensus between two board-certified pathologists, paired with slide-level ER, PR, and HER2 labels for each patient.
The authors benchmark six classification architectures (ResNet-50, DenseNet-121, DenseNet-121-Camelyon17, ConvNeXt-Tiny, RetCCL, and a frozen Virchow2 foundation model) and three instance segmentation models (CellViT++, Cerberus, HoVer-NeXt), and add multi-task learning baselines that jointly predict nucleus segmentation and biomarker status under a staged training protocol. Code and dataset are released via an anonymised open-science link.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper fills a specific gap in computational pathology. Most existing resources provide either per-nucleus annotations, but not both on the same patients.
BC-MultiSet is, as far as I can tell, the first H&E dataset that pairs consensus nucleus instance annotations for tumor, fibroblast, and inflammatory cells with ER, PR, and HER2 status on a single breast cancer cohort.
That combination is genuinely new and enables a class of morphology-to-phenotype studies that were not feasible on existing resources.
The annotation protocol is good. Two board-certified pathologists independently annotated 109,180 nuclei in QuPath and resolved disagreement by consensus. At that scale, manual annotation is expensive, and the authors deserve credit for the effort involved.
The evaluation design avoids the most common leakage mistake. Patient-level stratified 4-fold CV prevents patches from the same slide from crossing folds, and the stratification preserves class distribution for both biomarker and cell-type tasks.
The benchmark set is reasonably broad. Six classification architectures spanning ResNet, DenseNet, ConvNeXt, RetCCL, and Virchow2 cover the CNN, ViT, and foundation-model paradigms. Three segmentation baselines (CellViT++, Cerberus, HoVer-NeXt) represent the main families of modern nucleus instance segmentation.
The MTL analysis is honest. The paper explicitly documents negative transfer and introduces a staged training protocol (MTL-A versus MTL-B) that partially recovers it. It is uncommon for dataset papers to run MTL at all, and the framing is candid rather than oversold.
Code and data are released via an anonymised open-science link during review, which puts the submission above MICCAI’s minimum reproducibility bar.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The cohort is very small for the clinical claims being made. 50 patients with 4-fold patient-level CV leaves around 12 patients per test fold. Of those, roughly 4 are HER2-positive given the 16/34 class split. Any per-biomarker AUC estimate at that scale is dominated by sampling variance, and no repeated-CV or bootstrap CI is reported.
Standard deviations in Table 2 frequently exceed the effect size. RetCCL ER F1 is 53.3 ± 37.7, DenseNet121-C ER F1 is 66.8 ± 32.7, ResNet50 MOLT F1 is 45.7 ± 35.8, Virchow2 MOLT F1 is 35.1 ± 25.5.At least one fold is clearly collapsing for several models, and the paper does not investigate why.
One result in Table 2 is not internally consistent. HoVer-NeXt (A) and (B) report ER AUC of 47.2% (below chance) with ER F1 of 93.1%. An AUC near 0.5 indicates no ranking signal. With 70% ER prevalence, an always-positive predictor gives precision 0.70, recall 1.00, and therefore F1 ≈ 0.82, not 0.93.F1 of 0.93 combined with AUC of 0.47 is not reachable by any standard decision rule on these labels. This looks like either a metric-computation bug, threshold selection leaking from the test fold, or an undocumented F1 definition. As written, the two numbers cannot both be right, and the issue is not flagged.
MOLT appears to be collinear with ER. The cohort is reported as “35 luminal and 15 non-luminal,” which matches exactly the 35/15 ER-positive/negative split. If MOLT is operationally defined as ER status, it is the same task as ER and should not be reported as a separate target.
Slide-level biomarker labels are inherited by every 256×256 patch at 40x (~64 μm per side). Every patch, including pure stroma or pure normal-looking glands, carries the same ER/PR/HER2 label as the slide. This makes patch-level AUC a weak proxy for what the clinical task actually is. A slide-level MIL baseline (e.g. ABMIL, CLAM, TransMIL) is the correct comparison, and the paper explicitly defers this to future work, which is the wrong choice for a benchmark paper.
Table 4 reports a Cerberus fibroblast Dice of 0.8 ± 0.3, while every other model sits between 25 and 38.That is a 30-40x discrepancy. Either Cerberus is catastrophically failing on the fibroblast class and the paper does not say so, or this is a decimal-point typo. Either way it needs to be fixed or flagged in text.
Only three cell classes are annotated. PanNuke has five (neoplastic, inflammatory, connective, dead, epithelial), NuCLS has six. Merging cancer-associated fibroblasts with general stromal cells under “fibroblast” is also a coarser grouping than the CAF literature would want. The choice is defensible but is not justified in the paper.
Inter-rater agreement before consensus is not reported. For a dataset paper whose credibility rests on annotation quality, per-class Cohen’s kappa or instance-level IoU between the two pathologists is the standard number to report, and it is absent.
No external validation. TCGA-BRCA is listed in Table 1 for comparison but never used as a transferability test. A small external evaluation on TCGA-BRCA ER/PR/HER2 (or on the BACH / BCSS cohorts) would significantly strengthen the “benchmark” framing.
Cohort descriptors are thin. Patient age, tumor grade, stage, scanner/vendor, and per-patient patch counts are not reported. 5,586 patches across 50 patients averages 112 per patient, but imbalance is not addressed.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The data contribution is real and essential in the community. The concern is the benchmark, not the dataset. Several Table 2 results have standard deviations larger than effect sizes, at least one row (HoVer-NeXt ER AUC 47.2 / F1 93.1) is not internally consistent, the Cerberus fibroblast Dice of 0.8 looks like a decimal-point typo, and MOLT appears to be a relabelling of ER. The Virchow2 comparison is also asymmetric by construction, so the “foundation models win on classification” claim is confounded with a protocol choice. Because the benchmark is a core part of what the paper is selling (Section 4 is the bulk of the contribution), these issues matter.
None of the problems are terminal. The dataset itself is solid, the MTL framing is honest, and code plus data are released. If the rebuttal resolves the AUC/F1 anomaly, fixes the fibroblast Dice number, clarifies MOLT, and either re-runs a fair Virchow2 comparison or qualifies the claim, I would raise to Weak Accept. Adding inter-rater kappa and patient-level bootstrap CIs would cement the move.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank all reviewers for their constructive feedback and address each concern below.
1) [R3/AC] HoVer-NeXt ER AUC/F1 inconsistency: AUC was computed at the patient-level, while F1 was erroneously performed in the table at the patch-level for HoVer-NeXt only. Corrected patient-level values (ER AUC=55.6%, F1=78.6%) are internally consistent. Table 2 will be corrected.
2) [R3/AC] Cerberus fibroblast Dice=0.8%: This was a typographical error, but not in the Dice scores. The mistake concerns the PQ values: the correct results are 1.88±0.05 for Cerberus (not 18.8±0.05) and 1.97±0.5 for Cerberus+MTL (not 19.7±0.5). The gap versus CellViT++/HoVer-NeXt is architectural: Cerberus lacks explicit instance separation and specialized post-processing. This will be flagged explicitly in the revision.
3) [R3/AC] MOLT collinearity with ER: MOLT is defined via the St. Gallen criteria, which combine ER, PR, and HER2, not ER alone (Vasconcelos, The St. Gallen surr. classif. 2016). To eliminate confusion, we will rename this target “non-Luminal,” reflecting the challenge of identifying Triple-Negative/HER2-enriched subtypes.
4) [R3/AC] Asymmetric Virchow2 evaluation: Using a frozen encoder with linear probing is the standard protocol for billion-parameter foundation models to prevent catastrophic overfitting on small cohorts (Vorontsov, A found. mod. 2024; Kumar, Fine-tun. can dist. 2022). This will be stated explicitly.
5) [R3/AC] Large standard deviations: High variance reflects class imbalance which is confirmed by our experiments. This is not an experimental error. We will add per-fold breakdowns and bootstrapped 95% CIs.
6) [R1/R3/AC] No external validation; coarser ontology: External comparison is structurally impossible: TCGA-BRCA lacks per-nucleus annotations, and PanNuke/NuCLS lack receptor-status labels. Providing their combination is BC-MultiSet’s primary contribution. The three-class ontology maps to prognostically established TME components (Salgado, The eval. of TILs 2015). These limitations will be stated.
7) [R1/R3/AC] Small cohort; thin descriptors: We will add the following clinical details: median age 56 (IQR 45–62), grade I=6/II=28/III=16, stage I/II=33, III/IV=17.Our 50 WSIs with 109,180 annotated nuclei are consistent with established benchmarks: MoNuSeg (30 WSIs) and CoNSeP (41 images).
8) [R3/AC] Missing inter-rater agreement: At the instance level (IoU=0.7), agreement confirms annotation reliability: Tumor κ=0.83, Inflammatory κ=0.76, Fibroblast κ=0.60 (McHugh, Interrater reliability 2012). This will be included.
9) [R3/AC] Patch-level labels without MIL baselines: Our design emulates clinical HPF review: 10 patches (1000×1000 px) per WSI were extracted blind to IHC status. Baselines like ABMIL, CLAM, TransMIL require thousands of patches (Ilse, Att-based deep mult. 2018; Shao, Transmil: Transf. based cor. 2021), which in our cohort would cause severe overfitting. Label noise under blind extraction reflects real HPF-level uncertainty, representing the core challenge our benchmark targets. This rationale will be added to the final version.
10) [R2] Breast cancer motivation; annotation time: Breast cancer was selected because it is the most prevalent female malignancy; a single H&E slide captures tumor, stroma, and immune compartments simultaneously; and ER/PR/HER2 are IHC-validated labels linked to morphology. Median annotation time was 9 hours/WSI (via QuPath logs), totaling ~450 pathologist-hours.
11) [R2] Figure 2 legend: A color legend and in-text figure references will be added.
12) [R2] Methodological innovation: BC-MultiSet’s novelty lies in being the first dataset to pair per-nucleus polygon annotations with IHC-validated receptor-status labels. Furthermore, it features multi-task joint optimization across three architectures and introduces a staged MTL protocol (Strategy A/B) that mitigates negative transfer under limited data.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper introduces BC-MultiSet, a breast cancer dataset that uniquely pairs pixel-level instance segmentation masks for tumor, inflammatory and fibroblast nuclei with slide-level ER/PR/HER2 biomarkers from 50 invasive breast carcinoma patients. Reviewers appreciated the unique pairing of nucleus-level annotations with molecular biomarkers, the careful consensus annotation by two pathologists, the breadth of architecture comparison and the public release of code and data via anonymized links. However, the reviewers identified weaknesses including a small patient cohort (50) limiting patient-level analyses and biomarker stratification [R1, R3], evaluation only on the proposed dataset without external validation on TCGA, PanNuke or NuCLS [R1], absence of methodological innovation since the experiments use existing architectures [R2], missing motivation for the breast-cancer focus and time required for annotation [R2], very large standard deviations frequently exceeding effect sizes [R3], internally inconsistent results [R3], MOLT operationally collinear with ER status and reported as a separate target [R3], patch-level biomarker labels propagated from slides without slide-level MIL baselines [R3], coarser ontology than PanNuke/NuCLS without justification [R3] and missing inter-rater agreement statistics [R3]. The AC read the paper and the reviews and considers that the dataset itself is a valuable contribution but several benchmark integrity issues need to be clarified or fixed in a rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper received mixed reviews, and it was invited for rebuttal. The authors submitted a rebuttal to address the weaknesses raised by the reviewers. After the rebuttal, the two reviewers maintained their positive assessment and unfortunately the third reviewer did not update their score. The AC read the paper, reviews, rebuttal, and final justification, and agrees that the paper has merit and it would be interesting for the community. The AC advises the authors to update their paper according to the comments of the reviewers for the camera-ready version.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
This paper primarily presents a dataset and benchmarking study. For the dataset component, important details regarding annotation procedures, population diversity, representativeness, and data collection are difficult to fully assess given the limited scale (50 WSIs from a single site) and the constraints of the MICCAI 8-page format. In addition, the double-blind review process further limits the ability to evaluate aspects such as institutional expertise, acquisition protocols, and ethical oversight. As a result, this type of contribution appears more suitable for a journal setting than for MICCAI.
For the benchmarking component, concerns regarding experimental rigor remain insufficiently addressed after the rebuttal, and the claimed clinical value is still not fully convincing. Overall, this paper is ranked significantly below the 50th percentile of the rebuttal-stage papers in my batch.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The manuscript introduces a breast cancer histopathology image dataset and evaluates a number of deep learning models on the dataset for nuclei segmentation and classification. This dataset will be a valuable contribution to the medical image analysis and computer vision communities. The rebuttal addresses most of the reviewers’ concerns, such as unclear motivation for breast cancer focus, missing dataset characteristics, inconsistent experimental results and large standard deviations.