MHIM-MIL: Masked Hard Instance Mining

• The paper introduces MHIM-MIL, a framework that suppresses dominant evidence via masking to force models to utilize alternative discriminative cues.
• MHIM-MIL employs attention-based aggregation and teacher–student mechanisms, adapting to domains such as histopathology and clinical text for improved performance.
• Empirical evaluations show that MHIM-MIL increases robustness and explainability by reducing overfitting to salient instances, boosting metrics like AUC and RMSE.

Searching arXiv for MHIM-MIL and closely related work to ground the article in current literature. Masked Hard Instance Mining in Multiple Instance Learning (MHIM-MIL) denotes a class of weakly supervised learning procedures in which a model deliberately suppresses highly salient, typically easy-to-classify instances during training so that a learner must exploit less dominant but still discriminative evidence at the bag level. In the literature, the term is used across several related but non-identical formulations: it was introduced for whole-slide image (WSI) classification through a Siamese teacher–student framework with attention-based masking and consistency learning (Tang et al., 2023), extended for gigapixel histopathology with class-aware instance probabilities, large-scale random masking, and a Global Recycle Network (Tang et al., 15 Sep 2025), and adapted to text-based depression detection as a two-phase donor–receiver attention-masking strategy within a Multiple Instance Learning (MIL) setting (Prakrankamanant et al., 30 May 2025).

1. Conceptual basis and MIL formulation

MHIM-MIL is defined on the standard MIL assumption that a sample is a bag of instances with only bag-level supervision. In WSI analysis, a gigapixel slide is partitioned into image patches $X=\{x_i\}_{i=1}^N$, instance labels are unknown, and only the slide label $Y$ is provided (Tang et al., 2023). In depression detection, each interview is treated as a bag of sentences or dialogue parts, with labels available at the interview level rather than for individual instances; the Thai-Maywe dataset additionally provides Importance Sentence Labels (ISL) only for HAM-D 1 and only for evaluation, not training (Prakrankamanant et al., 30 May 2025).

The motivation is consistent across domains. Attention-based MIL methods tend to concentrate on a small subset of salient instances. In histopathology, these are often easy tissue regions that dominate optimization while harder regions near decision boundaries remain under-explored (Tang et al., 2023). In low-resource clinical text, attention can overfit to a few conspicuous cues, such as a mention of “suicide,” while neglecting other clinically relevant signals such as sleep disturbance or social withdrawal (Prakrankamanant et al., 30 May 2025). MHIM-MIL addresses this by masking or suppressing instances that the model currently finds easiest, thereby forcing broader evidence aggregation.

Paper	Domain	Defining MHIM-MIL mechanism
(Tang et al., 2023)	WSI classification	Attention-score masking with teacher–student EMA and consistency loss
(Tang et al., 15 Sep 2025)	Gigapixel histopathology	Class-aware probability masking, RHSM/RSM, Global Recycle Network
(Prakrankamanant et al., 30 May 2025)	Depression detection	Donor–receiver Dual Encoder with attention masking under regression
(Li et al., 2019)	Histopathology precursor	Attention-based hard negative mining without explicit masking

2. Core mechanism and mathematical structure

The common substrate is attention-based bag aggregation. For instance features $z_i$ or $h_i$, the bag representation is formed as

$F=\sum_{i=1}^{N} a_i z_i$

or, in the depression setting,

$z=\sum_i a_i h_i,$

with attention weights derived from logits such as

$$e_i=u^T\tanh(Wh_i), \qquad a_i=\frac{\exp(e_i)}{\sum_j \exp(e_j)}.$$

The bag-level predictor then maps the aggregated representation to either a class label or a continuous score (Prakrankamanant et al., 30 May 2025).

In the original WSI formulation, the teacher processes the full bag, computes attention scores, and produces a binary mask over instances. High Attention Masking (HAM) removes the top $\beta_h\%$ instances, Low Attention Masking (L-HAM) removes the bottom $\beta_l\%$, Random Masking (R-HAM) removes a random $\beta_r\%$, and hybrid masks combine these sets by union (Tang et al., 2023). The masked bag $Y$0 is then passed to the student, which learns from the residual, relatively harder subset. The teacher is updated by exponential moving average,

$Y$1

with $Y$2, and a consistency term aligns student and teacher bag representations (Tang et al., 2023).

The 2025 histopathology formulation changes the mining signal. Instead of relying only on class-agnostic attention, it computes class-aware instance probabilities from teacher features:

$Y$3

with binary or multi-class probabilities derived from attention-weighted instance features. The highest-probability instances are treated as easy and masked via Randomly High Score Masking (RHSM): the top $Y$4 instances are candidates, and half are randomly masked. A second masking stage then removes a large fraction of the remaining sequence either by Low Score Masking (LSM) or, more effectively in reported experiments, Random Score Masking (RSM) (Tang et al., 15 Sep 2025).

The depression-detection variant uses a two-phase donor–receiver procedure rather than EMA-based co-training. A donor Dual Encoder is first trained normally. In the second phase, donor attention on the same bag identifies top-attended instances; a random subset of those top-$Y$5 positions and an additional random subset from lower-attended positions are masked, with total masks $Y$6 and $Y$7 in experiments. The receiver then applies either exact masked softmax,

$Y$8

or masked embeddings $Y$9, and is trained with mean squared error on the bag-level regression target (Prakrankamanant et al., 30 May 2025).

3. Architectural variants and training regimes

In the original MHIM-MIL framework for WSI classification, the student can be any attention-based MIL model, including ABMIL, TransMIL, or DSMIL. ABMIL uses a gated attention mechanism, while TransMIL employs Transformer self-attention with a class token. For TransMIL specifically, the reported implementation uses first-layer attention rather than last-layer attention for mining hard instances, and multi-head attention is fused by a voting strategy rather than naive averaging (Tang et al., 2023).

The later gigapixel histopathology version retains the teacher–student Siamese structure but adds a Global Recycle Network (GRN). After large-ratio masking, the masked subset $z_i$0 is not discarded outright; instead, multi-head cross-attention between learned global queries $z_i$1 and $z_i$2 produces recycled features

$z_i$3

and the global queries are updated by EMA,

$z_i$4

The student receives the concatenated hard input $z_i$5, combining unmasked hard instances with globally summarized information from masked instances (Tang et al., 15 Sep 2025). This addition is explicitly intended to mitigate the loss of critical features caused by aggressive random masking.

The depression architecture differs both in backbone and supervision. Its backbone is a Dual Encoder consisting of a Prefix Encoder and a Sentence Encoder, followed by a Bi-LSTM, a single attention layer, and a regression head. English models use RoBERTa-base with all-mpnet-base-v2; Thai models use WangchanBERTa with ConGen. The task targets severity scores such as HAM-D 1, overall HAM-D, or PHQ-8 item 2, and the optimization objective is MSE rather than cross-entropy (Prakrankamanant et al., 30 May 2025). Inference uses only the trained receiver, without masking and without the donor.

These variants show that MHIM-MIL is not tied to a single backbone, a single prediction type, or a single masking criterion. What remains stable is the training principle: an auxiliary model or auxiliary phase identifies instances whose current salience is so dominant that suppressing them may reveal broader discriminative structure.

4. Objectives, optimization, and computational properties

The original WSI formulation optimizes a bag-level classification loss together with a teacher–student consistency loss. With binary prediction $z_i$6, the classification term is

$z_i$7

and the consistency loss is

$z_i$8

where $z_i$9 and $h_i$0. The total objective is

$h_i$1

with $h_i$2 reported in the 2023 paper and $h_i$3 tuned by validation (Tang et al., 2023).

The 2025 histopathology version preserves the same general form,

$h_i$4

but uses class-aware instance probability for mask construction and cross-entropy over student logits in experiments. Optimization uses Adam with learning rate $h_i$5, weight decay $h_i$6, cosine annealing, batch size 1, and early stopping with patience 30 on CAMELYON and 20 on TCGA for diagnosis and subtyping; survival analysis is trained for 30 epochs (Tang et al., 15 Sep 2025).

The depression variant is structurally simpler. Phase 1 trains the donor with

$h_i$7

and Phase 2 trains the receiver on masked bags with

$h_i$8

No explicit regularizers such as entropy penalties or KL terms are reported. Training uses AdamW, learning rate $h_i$9, 10% warmup, linear decay, start and end at $F=\sum_{i=1}^{N} a_i z_i$0, and 200 epochs (Prakrankamanant et al., 30 May 2025).

Computationally, MHIM-MIL adds auxiliary cost because masking requires teacher-side scoring and, in some variants, sorting instances within each bag. In the depression setting, per-bag masking requires sorting donor attention in $F=\sum_{i=1}^{N} a_i z_i$1 and training two models increases time and memory relative to baseline Dual Encoder training (Prakrankamanant et al., 30 May 2025). In the histopathology setting, the effect depends on backbone. For TransMIL on CAMELYON, the 2025 version reports 13.2 s / 10.6 GB for vanilla TransMIL and 10.5 s / 5.6 GB for MHIM-v2(TransMIL), corresponding to a $F=\sum_{i=1}^{N} a_i z_i$2 reduction in training time and a $F=\sum_{i=1}^{N} a_i z_i$3 reduction in memory because masking shortens the effective sequence length (Tang et al., 15 Sep 2025).

5. Empirical performance and evaluation methodology

Evaluation differs by task. In depression detection, prediction quality is measured by RMSE and MAE, while explainability is quantified by attention entropy, Recall@k against ISL, and deletion sensitivity under removal of top-attended sentences (Prakrankamanant et al., 30 May 2025). In histopathology diagnosis and subtyping, the primary metrics are AUC, Accuracy, and F1-score; in survival analysis, the metric is C-index (Tang et al., 15 Sep 2025).

On Thai-Maywe HAM-D 1, the Dual Encoder baseline reports RMSE $F=\sum_{i=1}^{N} a_i z_i$4 and MAE $F=\sum_{i=1}^{N} a_i z_i$5, while MHIM reports RMSE $F=\sum_{i=1}^{N} a_i z_i$6 and MAE $F=\sum_{i=1}^{N} a_i z_i$7; attention entropy increases from $F=\sum_{i=1}^{N} a_i z_i$8 to $F=\sum_{i=1}^{N} a_i z_i$9. For overall HAM-D, the baseline reports RMSE $z=\sum_i a_i h_i,$0 and MAE $z=\sum_i a_i h_i,$1, while MHIM reports RMSE $z=\sum_i a_i h_i,$2 and MAE $z=\sum_i a_i h_i,$3, with entropy increasing from $z=\sum_i a_i h_i,$4 to $z=\sum_i a_i h_i,$5. On DAIC-WOZ for PHQ-8 item 2, RMSE decreases from $z=\sum_i a_i h_i,$6 to $z=\sum_i a_i h_i,$7, and deletion analysis indicates improved explainability. For HAM-D 1 Recall@10%, the Dual Encoder reports $z=\sum_i a_i h_i,$8 and MHIM reports $z=\sum_i a_i h_i,$9, with gains persisting across larger $$e_i=u^T\tanh(Wh_i), \qquad a_i=\frac{\exp(e_i)}{\sum_j \exp(e_j)}.$$0 values (Prakrankamanant et al., 30 May 2025).

The 2023 WSI paper reports strong gains on CAMELYON-16 and TCGA Lung Cancer. On CAMELYON-16, AUC improves from $$e_i=u^T\tanh(Wh_i), \qquad a_i=\frac{\exp(e_i)}{\sum_j \exp(e_j)}.$$1 for ABMIL, $$e_i=u^T\tanh(Wh_i), \qquad a_i=\frac{\exp(e_i)}{\sum_j \exp(e_j)}.$$2 for TransMIL, and $$e_i=u^T\tanh(Wh_i), \qquad a_i=\frac{\exp(e_i)}{\sum_j \exp(e_j)}.$$3 for DSMIL to $$e_i=u^T\tanh(Wh_i), \qquad a_i=\frac{\exp(e_i)}{\sum_j \exp(e_j)}.$$4 for MHIM-MIL(ABMIL), $$e_i=u^T\tanh(Wh_i), \qquad a_i=\frac{\exp(e_i)}{\sum_j \exp(e_j)}.$$5 for MHIM-MIL(TransMIL), and $$e_i=u^T\tanh(Wh_i), \qquad a_i=\frac{\exp(e_i)}{\sum_j \exp(e_j)}.$$6 for MHIM-MIL(DSMIL). On TCGA, AUC improves from $$e_i=u^T\tanh(Wh_i), \qquad a_i=\frac{\exp(e_i)}{\sum_j \exp(e_j)}.$$7 for ABMIL and $$e_i=u^T\tanh(Wh_i), \qquad a_i=\frac{\exp(e_i)}{\sum_j \exp(e_j)}.$$8 for TransMIL to $$e_i=u^T\tanh(Wh_i), \qquad a_i=\frac{\exp(e_i)}{\sum_j \exp(e_j)}.$$9 and $\beta_h\%$0, respectively; MHIM-MIL(DSMIL) reports $\beta_h\%$1 (Tang et al., 2023).

The 2025 gigapixel extension broadens the evidence base to 12 benchmarks covering diagnosis, subtyping, survival analysis, and cross-source validation. On CAMELYON with ResNet-50 features, MHIM-v2(TransMIL) reports AUC $\beta_h\%$2, compared with $\beta_h\%$3 for TransMIL, while MHIM-v2(ABMIL) reports $\beta_h\%$4 versus $\beta_h\%$5 for ABMIL. On TCGA-NSCLC with ResNet-50, MHIM-v2(DSMIL) reports $\beta_h\%$6 and MHIM-v2(TransMIL) reports $\beta_h\%$7. In survival analysis with UNI features, MHIM-v2(TransMIL) reports C-index $\beta_h\%$8 for LUAD, $\beta_h\%$9 for BLCA, and $\beta_l\%$0 for LUSC (Tang et al., 15 Sep 2025).

A notable empirical pattern is that gains are larger in low-resource or large-bag settings where saliency bias is especially pronounced. This is explicit in the depression study, which reports more pronounced improvements in Thai than in English, and in the histopathology studies, where Transformer backbones benefit both in accuracy and in sequence-length reduction through masking (Prakrankamanant et al., 30 May 2025).

6. Explainability, antecedents, and limitations

Explainability is a central theme in the depression adaptation and an implicit one in WSI analysis. In depression detection, attention weights are treated as interpretable rationales. Qualitative examples report that MHIM shifts top-ranked sentences toward broader symptom mentions such as “Lately, I don't really feel like meeting anyone” and “I've been feeling sad from time to time,” whereas baseline selections may include less informative statements such as “I eat normally” (Prakrankamanant et al., 30 May 2025). In histopathology, visualizations show reduced false positives in lightly stained adipose regions, stronger emphasis on lymph node margins, and improved alignment between high tumor-probability patches and pathologist annotations (Tang et al., 15 Sep 2025).

The immediate antecedent is not masked but closely related hard-instance mining. “Deep Instance-Level Hard Negative Mining Model for Histopathology Images” learns attention weights to identify key patches, introduces adaptive weighing to emphasize hard samples within a bag, and then generates hard negative bags from high-attention instances in false-positive negative slides (Li et al., 2019). That work does not explicitly mask instances during pooling, but it establishes two ideas later absorbed into MHIM-MIL research: first, that weakly supervised attention can serve as a proxy for instance difficulty; second, that training benefits when misleading or difficult instances are deliberately reintroduced into optimization.

Several limitations recur across the literature. Hyperparameter sensitivity is explicit: mask ratios such as $\beta_l\%$1, $\beta_l\%$2, $\beta_l\%$3, the consistency weight $\beta_l\%$4, temperature $\beta_l\%$5, and the depression parameters $\beta_l\%$6, $\beta_l\%$7, and $\beta_l\%$8 all require validation-based tuning (Tang et al., 2023). Overly aggressive masking can remove informative instances and degrade performance; this is why Randomly HAM, cosine decay of $\beta_l\%$9, and GRN-style recycling are introduced (Tang et al., 15 Sep 2025). Teacher quality also matters: the WSI papers report that “Init.+Momentum” is more stable and better-performing than weaker teacher choices, and the depression paper notes that suppressing truly critical features may hurt performance if the masking ratio is too large (Tang et al., 15 Sep 2025).

A common misconception is that MHIM-MIL simply “finds the most important instances.” The operative mechanism is the opposite: it suppresses the most dominant evidence during training so that the learner must recruit alternative evidence. Another misconception is that attention scores are a universally adequate hardness proxy. The 2025 WSI formulation explicitly replaces attention-only masking with class-aware instance probability because attention is class-agnostic, while the depression formulation continues to rely on attention because instance-level labels are largely unavailable and the task is framed as weakly supervised MIL (Tang et al., 15 Sep 2025).

Reported future directions are correspondingly pragmatic rather than doctrinal. They include softer masks based on logit penalties rather than zeroing, entropy or KL regularization for stability, multi-head masking and head-wise diversification, integration with gradient-based attributions, broader datasets, multimodal inputs, and more robust MIL assumptions when a single highly relevant sentence can induce false positives (Prakrankamanant et al., 30 May 2025). The histopathology literature additionally suggests extension to other large-bag, weakly supervised domains such as audio event detection, video anomaly detection, and remote sensing, where sparse positives and saliency collapse have similar structural effects (Tang et al., 15 Sep 2025).