ContractionPPO: An Unverified MIL Technique

• ContractionPPO is an unresolved term in current literature, with no primary source formally defining its architecture or objective.
• It is often conflated with masked hard instance mining methods used in gigapixel histopathology image analysis and text-based depression detection.
• The available research documents rigorous MIL frameworks with diverse masking techniques and optimization protocols, leaving ContractionPPO as a placeholder pending formal definition.

Searching arXiv for "ContractionPPO" and closely related terms to ground the article in published work. ContractionPPO is not defined in the supplied source set. The cited arXiv records instead concern masked hard instance mining in multiple instance learning, particularly for gigapixel histopathology and text-based depression detection, together with earlier instance-level hard negative mining for histopathology (Tang et al., 15 Sep 2025). In the absence of a primary source explicitly introducing or naming ContractionPPO, the term cannot be given a rigorous method definition, objective function, algorithmic workflow, or benchmark profile without moving beyond the evidentiary basis. A plausible implication is that ContractionPPO refers either to a work not included in the present corpus or to a naming variant not represented in the supplied records.

1. Attestation status

Within the provided material, no title, abstract, or methodological description uses the name ContractionPPO. The available papers are "Multiple Instance Learning Framework with Masked Hard Instance Mining for Gigapixel Histopathology Image Analysis" (Tang et al., 15 Sep 2025), "Explainable Depression Detection using Masked Hard Instance Mining" (Prakrankamanant et al., 30 May 2025), "Multiple Instance Learning Framework with Masked Hard Instance Mining for Whole Slide Image Classification" (Tang et al., 2023), and "Deep Instance-Level Hard Negative Mining Model for Histopathology Images" (Li et al., 2019).

This absence is methodologically important. For an encyclopedia treatment aimed at arXiv-reading researchers, a named method ordinarily requires at minimum a stable expansion of the acronym, a formal objective, an architectural specification, an optimization protocol, and an empirical evaluation regime. None of those elements is available here for ContractionPPO itself. Any attempt to supply them directly would therefore be inferential rather than documentary.

2. What the supplied literature actually covers

The dominant theme in the provided corpus is masked hard instance mining under weak supervision. In the histopathology setting, a gigapixel whole-slide image is partitioned into non-overlapping patches, forming a bag $X = \{x_i\}_{i=1}^N$ with slide-level label $Y$, while instance labels are unknown. Embedding-level MIL maps each instance to a feature $z_i \in \mathbb{R}^D$, aggregates those features into a bag representation $F$, and predicts the bag label. The 2025 MHIM-MIL paper states that attention-based MIL tends to bias learning toward easy-to-classify salient instances, and proposes masked hard instance mining to suppress easy instances and force learning from harder ones (Tang et al., 15 Sep 2025).

The 2023 conference version presents the same broad idea for WSI classification through a teacher–student Siamese structure with a consistency constraint and EMA teacher updates, while the 2025 extension elaborates class-aware instance probability, randomly high score masking, large-scale random masking, and a global recycle network. The text-based depression detection paper adapts MHIM to a dialogue-level MIL formulation, where instances are question–response pairs or utterances, attention weights provide explanations, and the training objective is Mean Squared Error on bag-level severity scores. The 2019 histopathology paper differs from explicit masking: it uses attention, adaptive weighing, and hard negative bag generation rather than a masked teacher–student framework (Prakrankamanant et al., 30 May 2025).

3. Core methodological motifs in the adjacent literature

The most detailed adjacent framework is MHIM-MIL. Its student model processes a mined subset of instances, while a momentum teacher with the same architecture evaluates full-bag evidence, generates masks, and is updated by exponential moving average. In the 2025 formulation, the teacher computes class-aware instance probabilities via $S = C_T(A \cdot Z)$, where $A = [a_1,\dots,a_N] = T(Z)$ and $(A \cdot Z)_i = a_i z_i$. For binary tasks, a typical form is $p_i = \sigma(w^T(a_i z_i) + b)$; for multi-class tasks, $p_i^c = \mathrm{softmax}(W(a_i z_i) + b)_c$. The overall training objective is $L = L_{\mathrm{cls}} + \alpha L_{\mathrm{con}}$, with $Y$0 defined from softened teacher targets and student bag embeddings, and the teacher update is $Y$1 with $Y$2 in experiments (Tang et al., 15 Sep 2025).

Mask construction is central. The teacher sorts instances by class-aware probability, then applies Randomly High Score Masking, where the top $Y$3 instances are taken as candidates and half are randomly masked; a cosine decay schedule reduces $Y$4 over training. After that, large-scale random masking with $Y$5–$Y$6 reduces redundancy and sequence length. To recover dropped global signals, the global recycle network reconstructs features from the masked subset via multi-head cross-attention using learnable queries $Y$7, followed by an EMA-style query update. The 2023 version instead organizes masking as combinations of high-attention masking, low-attention masking, random masking, and hybrid unions, again under an EMA teacher and consistency loss (Tang et al., 2023).

These motifs are specific to MHIM-style MIL and should not be silently transferred to ContractionPPO. The only justified statement is that the supplied corpus contains sophisticated masking-based weak-supervision methods, not a documented ContractionPPO algorithm.

4. Experimental regimes represented in the source corpus

The empirical scope of the supplied literature is likewise well defined, but it concerns MHIM-family methods rather than ContractionPPO. The 2025 histopathology paper evaluates diagnosis, subtyping, survival analysis, and cross-source validation. Diagnosis and subtyping use AUC, Accuracy, and F1-score; survival uses C-index. Datasets include CAMELYON-16/17 merged for binary metastasis diagnosis, TCGA-NSCLC and TCGA-BRCA for subtyping, and TCGA-LUAD, TCGA-LUSC, and TCGA-BLCA for survival analysis. The paper reports that MHIM-v2 outperforms recent baselines across 12 benchmarks and gives specific efficiency comparisons for TransMIL, including reductions from $Y$8 and $Y$9 memory to $z_i \in \mathbb{R}^D$0 and $z_i \in \mathbb{R}^D$1 memory under MHIM-v2(TransMIL) on selected settings (Tang et al., 15 Sep 2025).

The depression-detection adaptation uses Thai-Maywe and DAIC-WOZ, formulates the task as bag-level regression, and evaluates RMSE, MAE, Recall@k against human Importance Sentence Labels, attention entropy, and deletion sensitivity. The model architecture uses a Dual Encoder, Bi-LSTM aggregation, and additive attention; MHIM is applied in a two-phase donor–receiver training pipeline in which high-attention and random instances are masked during receiver training. Reported improvements include RMSE $z_i \in \mathbb{R}^D$2 versus $z_i \in \mathbb{R}^D$3 on Thai HAM-D 1 for Dual Encoder + MHIM versus Dual Encoder, and a slight RMSE improvement on DAIC-WOZ, $z_i \in \mathbb{R}^D$4 versus $z_i \in \mathbb{R}^D$5 (Prakrankamanant et al., 30 May 2025).

The 2019 histopathology work addresses colon and breast cancer histopathology with Accuracy, Precision, Recall, F-score, AUC, and False Positive Rate, and reports gains from hard negative bag generation, especially with feature-clustered multiple bags. These results establish a lineage of instance-level hardness modeling in MIL, but not a ContractionPPO benchmark tradition (Li et al., 2019).

5. Conceptual boundaries and likely misconceptions

A common mistake would be to equate any hard-instance-mining or teacher–student masking method with ContractionPPO merely because the term is unavailable in the supplied sources. The literature provided here is narrowly centered on MIL under weak supervision, particularly in histopathology and clinical text. Its core operators are instance attention, masking policies over instances, bag-level aggregation, consistency regularization, and EMA stabilization. The papers explicitly situate themselves relative to attention selection, co-teaching, hard example mining, and mean-teacher dynamics, not under the name ContractionPPO (Tang et al., 15 Sep 2025).

Another likely misconception is that the 2025 MHIM-MIL paper and the 2023 conference paper define distinct named families unrelated to one another. The record explicitly states that MHIM-v2 is the improved version presented in the 2025 paper, while the 2023 paper is the ICCV conference version. The extension adds class-aware instance probability, large-scale random masking, and the global recycle network, whereas the earlier version formulates masking more directly from teacher attention scores (Tang et al., 15 Sep 2025).

A further boundary is terminological. Because the suffix “PPO” is not expanded anywhere in the supplied evidence, any statement that ContractionPPO belongs to a specific optimization family would be speculative. The only defensible editorial position is to withhold such identification pending a primary source.

6. Requirements for a proper technical entry

A rigorous encyclopedia entry on ContractionPPO would require a primary reference that establishes at least five elements: the exact name and acronym expansion; the problem setting; the formal update rule or loss; implementation details sufficient to distinguish it from adjacent methods; and an evaluation protocol with explicit baselines and metrics. The supplied papers illustrate what such documentation looks like in practice. They specify bag formation, attention formulas, masking operators, optimization schedules, EMA updates, hyperparameters, datasets, and reported outcomes in a level of detail that supports reproduction (Tang et al., 2023).

Until comparable documentation is available for ContractionPPO, the term is best treated as an unresolved bibliographic entry rather than a settled method. The present corpus supports a detailed account of masked hard instance mining in MIL, but it does not support attributing those mechanisms, formulas, or results to ContractionPPO. A plausible implication is that future disambiguation would require locating a missing arXiv record or a differently titled primary paper in which the method is explicitly introduced.