Pancreatic Cancer ROSE Image Classification Based on Multiple Instance Learning with Shuffle Instances

Tianyi Zhang, Youdan Feng, Yu Zhao, Yunlu Feng, Yanli Lei, Nan Ying, Fan Song, Zhiling Yan, Yufang He, Aiming Yang, and Guanglei Zhang, “Shuffle Instances-based Vision Transformer for Pancreatic Cancer ROSE Image Classification” Computer Methods and Programs in Biomedicine, 244, p. 107969. (SCI, Q1, IF=6.1)

https://authors.elsevier.com/a/1iDRxcV4LHEkF

• Results can be view in Archive folder

• Colab scripts are provided with a sample dataset (MICCAI 2015 chanllange)

Abstract

Background and Objective: The rapid on-site evaluation (ROSE) technique improves pancreatic cancer diagnosis by enabling immediate analysis of fast-stained cytopathological images. Automating ROSE classification could not only reduce the burden on pathologists but also broaden the application of this increasingly popular technique. However, this approach faces substantial challenges due to complex perturbations in color distribution, brightness, and contrast, which are influenced by various staining environments and devices. Additionally, the pronounced variability in cancerous patterns across samples further complicates classification, underscoring the difficulty in precisely identifying local cells and establishing their global relationships.

Methods: To address these challenges, we propose an instance-aware approach that enhances the Vision Transformer with a novel shuffle instance strategy (SI-ViT). Our approach presents a shuffle step to generate bags of shuffled instances and corresponding bag-level soft-labels, allowing the model to understand relationships and distributions beyond the limited original distributions. Simultaneously, combined with an un-shuffle step, the traditional ViT can model the relationships corresponding to the sample labels. This dual-step approach helps the model to focus on inner-sample and cross-sample instance relationships, making it potent in extracting diverse image patterns and reducing complicated perturbations.

Results: Compared to state-of-the-art methods, significant improvements in ROSE classification have been achieved. Aiming for interpretability, equipped with instance shuffling, SI-ViT yields precise attention regions that identifying cancer and normal cells in various scenarios. Additionally, the approach shows excellent potential in pathological image analysis through generalization validation on other datasets.

Conclusions: By proposing instance relationship modeling through shuffling, we introduce a new insight in pathological image analysis. The significant improvements in ROSE classification leads to protential AI-on-site applications in pancreatic cancer diagnosis. The code and results are publicly available at https://github. com/sagizty/MIL-SI.

Method Overview

Overview of our proposed approach MIL-SI, composed of two steps MIL step and CLS step. In the data processing as illustrated in (a), the images will be transformed into patch-es, and the patch annotation label will be calculated based on the corresponding masks. In the MIL step, the bags of patches within a batch will be shuffled while the bags of image patches will remain unchanged in the CLS step. The bags are then composed into images with the soft-label aggregated from the patch-level label. In the 2-step training process in (b), after the feature extraction of the backbone, the patch tokens will be used to regress the bag-level soft-label in the MIL head. In the CLS step, an additional CLS head will be used to predict the categories of the input images based on the class token.

Results on the test set

MIL-SI

Model	Model info	MIL Info	size	Acc (%)	Precision (%)	Recall (%)	Sensitivity (%)	Specificity (%)	NPV (%)	F1_score (%)
ViT_384_401_PT_lf05_b4_p32_ROSE_MIL	MIL ViT	CLS+CLS_MIL+MIL	384, P32	94.00	91.98	90.68	90.68	95.77	95.05	91.32

SOTA models

Model	Model info	MIL Info	size	Acc (%)	Precision (%)	Recall (%)	Sensitivity (%)	Specificity (%)	NPV (%)	F1_score (%)
vgg16_384_401_PT_lf05_b4_ROSE_CLS	VGG 16	CLS	384	90.65	86.27	87.01	87.01	92.60	93.02	86.64
vgg19_384_401_PT_lf05_b4_ROSE_CLS	VGG 19	CLS	384	90.06	90.42	79.94	79.94	95.47	89.90	84.86
mobilenetv3_384_401_PT_lf05_b4_ROSE_CLS	Mobilenet v3	CLS	384	89.57	91.06	77.68	77.68	95.92	88.94	83.84
efficientnet_b3_384_401_PT_lf05_b4_ROSE_CLS	Efficientnet_b3	CLS	384	89.57	85.03	85.03	85.03	91.99	91.99	85.03
ResNet50_384_401_PT_lf05_b4_ROSE_CLS	ResNet50	CLS	384	90.75	87.36	85.88	85.88	93.35	92.51	86.61
inceptionv3_384_401_PT_lf05_b4_ROSE_CLS	Inception v3	CLS	384	90.75	86.72	86.72	86.72	92.90	92.90	86.72
xception_384_401_PT_lf05_b4_ROSE_CLS	Xception	CLS	384	90.94	91.46	81.64	81.64	95.92	90.71	86.27
swin_b_384_401_PT_lf05_b4_ROSE_CLS	Swin Transformer	CLS	384	89.17	86.75	81.36	81.36	93.35	90.35	83.97
ViT_384_401_PT_lf05_b4_ROSE_CLS	ViT	CLS	384	90.65	88.20	84.46	84.46	93.96	91.88	86.29
conformer_384_401_PT_lf05_b4_ROSE_CLS	Conformer	CLS	384	89.67	90.82	78.25	78.25	95.77	89.17	84.07
cross_former_224_401_PT_lf05_b4_ROSE_CLS	Cross_former	CLS	384	89.67	86.94	82.77	82.77	93.35	91.02	84.80
PC_Hybrid2_384_401_PT_lf05_b4_ROSE_CLS	MSHT	CLS	384	90.65	90.60	81.64	81.64	95.47	90.67	85.88

Counterpart augmentations

Model	Model info	MIL Info	size	Acc (%)	Precision (%)	Recall (%)	Sensitivity (%)	Specificity (%)	NPV (%)	F1_score (%)
ViT_384_401_PT_lf05_b4_ROSE_CutMix_CLS	ViT	CLS	384	92.72	89.55	89.55	89.55	94.41	94.41	89.55
ViT_384_401_PT_lf05_b4_ROSE_Cutout_CLS	ViT	CLS	384	92.32	91.07	86.44	86.44	95.47	92.94	88.70
ViT_384_401_PT_lf05_b4_ROSE_Mixup_CLS	ViT	CLS	384	92.52	88.83	89.83	89.83	93.96	94.53	89.33

Different head structure

Model	Model info	MIL Info	size	Acc (%)	Precision (%)	Recall (%)	Sensitivity (%)	Specificity (%)	NPV (%)	F1_score (%)
PC_ViT_384_401_PT_lf05_b4_ROSE_CLS	ViT	CLS	384	90.65	88.20	84.46	84.46	93.96	91.88	86.29
ViT_384_401_PT_lf05_b4_p32_NS_ROSE_MIL	MIL ViT (no shuffle MIL)	CLS+CLS_MIL	384, P32	92.13	90.06	87.01	87.01	94.86	93.18	88.51
ViT_384_401_PT_lf05_b4_p32_NCLSMIL_ROSE	MIL ViT	CLS+MIL, no cls step MIL regression	384, P32	93.41	91.59	89.27	89.27	95.62	94.34	90.41
ViT_384_401_PT_lf05_b4_p32_ROSE_MIL	MIL ViT	CLS+CLS_MIL+MIL	384, P32	94.00	91.98	90.68	90.68	95.77	95.05	91.32

Different patch size

Model	Model info	MIL Info	size	Acc (%)	Precision (%)	Recall (%)	Sensitivity (%)	Specificity (%)	NPV (%)	F1_score (%)
ViT_384_401_PT_lf05_b4_p16_ROSE_MIL	MIL ViT	CLS+CLS_MIL+MIL	384, P16	93.60	92.88	88.42	88.42	96.37	93.96	90.59
ViT_384_401_PT_lf05_b4_p32_ROSE_MIL	MIL ViT	CLS+CLS_MIL+MIL	384, P32	94.00	91.98	90.68	90.68	95.77	95.05	91.32
ViT_384_401_PT_lf05_b4_p64_ROSE_MIL	MIL ViT	CLS+CLS_MIL+MIL	384, P64	93.11	91.04	88.98	88.98	95.32	94.18	90.00
ViT_384_401_PT_lf05_b4_p128_ROSE_MIL	MIL ViT	CLS+CLS_MIL+MIL	384, P128	92.62	92.92	85.31	85.31	96.53	92.47	88.95

Different head balance

Model	Model info	MIL Info	size	Acc (%)	Precision (%)	Recall (%)	Sensitivity (%)	Specificity (%)	NPV (%)	F1_score (%)
ViT_384_401_PT_lf05_b4_p32_MIL_05_ROSE	MIL ViT	CLS+0.5CLS_MIL+0.5MIL	384, P32	91.73	84.62	93.22	93.22	90.94	96.17	88.71
ViT_384_401_PT_lf05_b4_p32_MIL_12_ROSE	MIL ViT	CLS+1.2CLS_MIL+1.2MIL	384, P32	92.52	87.77	91.24	91.24	93.20	95.22	89.47
ViT_384_401_PT_lf05_b4_p32_MIL_15_ROSE	MIL ViT	CLS+1.5CLS_MIL+1.5MIL	384, P32	93.31	91.81	88.70	88.70	95.77	94.07	90.23
ViT_384_401_PT_lf05_b4_p32_MIL_18_ROSE	MIL ViT	CLS+1.8CLS_MIL+1.8MIL	384, P32	93.41	91.12	89.83	89.83	95.32	94.60	90.47
ViT_384_401_PT_lf05_b4_p32_MIL_25_ROSE	MIL ViT	CLS+2.5CLS_MIL+2.5MIL	384, P32	93.50	92.35	88.70	88.70	96.07	94.08	90.49
ViT_384_401_PT_lf05_b4_p32_MIL_30_ROSE	MIL ViT	CLS+3.0CLS_MIL+3.0MIL	384, P32	93.60	91.40	90.11	90.11	95.47	94.75	90.75

Attention visuallization by grad-CAM

CAM on shuffled instances

bad caces

CAM of different patch settings

CAM of different settings on shuffled samples

More samples can be viewed in the folder of Archive