text stringlengths 0 5.79k |
|---|
### Prognostic Role of Patterns and Pattern‐Based Grading |
To evaluate the prognostic value of PATQUANT, we conducted a large‐scale validation using five well‐characterized patient cohorts with resectable LUAD (Figure 1D, detailed clinicopathological characteristics provided in Table 1) originating from three countries (Germany, Japan, the United States) and representing diffe... |
[TABLE] Clinicopathological characteristics of study cohorts. |
Parameter | TCGA | PLCO | NLST | UKK | KAMEDA |
Patients, n | 457 | 215 | 208 | 82 | 158 |
Whole slide images, n | 520 | 445 | 386 | 154 | 277 |
Whole slide image n > 1, cases | 21 (4.6%) | 166 (77.2%) | 151 (72.6%) | 50 (61.0%) | 63 (39.9%) |
Whole slide image n, range | 1–10 | 1–3 | 1–4 | 1–5 | 1–8 |
Sex |
Female | 247 (54.1%) | 112 (52.1%) | 102 (49%) | 41 (50.0%) | 76 (48.1%) |
Male | 219 (45.9%) | 103 (47.9%) | 106 (51%) | 41 (50.0%) | 82 (51.9%) |
Age |
Minimum | 33 | 52 | 55 | 44 | 42 |
Maximum | 88 | 74 | 74 | 87 | 85 |
Mean | 64.9 | 63.6 | 64.0 | 65.2 | 66.9 |
Median | 65.0 | 64.0 | 64.0 | 66.0 | 67.5 |
pT stage |
pT1 | 157 (34.4%) | 111 (51.6%) | 148 (71.2%) | 41 (50.0%) | 120 (75.9%) |
pT2 | 240 (52.5%) | 86 (40.0%) | 39 (18.8%) | 17 (20.7%) | 27 (17.1%) |
pT3 | 41 (9.0%) | 5 (2.3%) | 13 (6.3%) | 12 (14.6%) | 6 (3.8%) |
pT4 | 16 (3.5%) | 13 (6.0%) | 6 (2.9%) | 12 (14.6%) | 3 (1.9%) |
Unavailable | 3 (0.7%) | − | 2 (1.0%) | − | 2 (1.3%) |
pN stage |
pN0 | 300 (65.6%) | 169 (78.6%) | 151 (72.6%) | 50 (61.0%) | 133 (84.2%) |
pN1 | 89 (19.5%) | 23 (10.7%) | 24 (11.5%) | 14 (17.1%) | 15 (9.5%) |
pN2‐pN3 | 57 (12.5%) | 22 (10.2%) | 18 (8.7%) | 18 (22.0%) | 10 (6.3%) |
Unavailable | 11 (2.4%) | 1 (0.5%) | 15 (7.2%) | − | − |
UICC stage |
I | 249 (54.5%) | 147 (68.4%) | 135 (64.9%) | 31 (37.8%) | 119 (75.3%) |
II | 112 (24.5%) | 22 (10.2%) | 23 (11.1%) | 20 (24.4%) | 22 (13.9%) |
III | 63 (13.8%) | 30 (14.0%) | 24 (11.5%) | 26 (31.7%) | 13 (8.2%) |
IV | 25 (5.5%) | 16 (7.4%) | 8 (3.8%) | 5 (6.1%) | 2 (1.3%) |
Unavailable | 8 (1.8%) | − | 18 (8.7%) | − | 2 (1.3%) |
OS |
Alive | 289 (63.2%) | 52 (24.2%) | 98 (47.1%) | 43 (52.4%) | 124 78.5%) |
Dead | 156 (34.1%) | 163 (75.8%) | 110 (52.9%) | 39 (47.6%) | 34 21.5%) |
Unavailable | 12 (2.6%) | − | − | − | − |
CSS |
Non‐cancer deaths/alive | 315 (68.9%) | 110 (51.2%) | 36 (17.3%) | − | − |
Cancer death | 89 (19.5%) | 105 (48.8%) | 74 (35.6%) | − | − |
Unavailable | 53 (11.6%) | − | 98 (47.1%) | − | − |
PFS |
No progression | 252 (55.1%) | − | 132 (63.5%) | 48 (58.5%) | − |
Progression | 116 (25.4%) | − | 52 (25.0%) | 22 (26.8%) | − |
Unavailable | 89 (19.5%) | − | 24 (11.5%) | 12 (14.6%) | − |
Follow‐up duration |
Range in months | 1–242 | 1–278 | 1–159 | 2–108 | 2–254 |
Mean | 31.5 | 92.7 | 94.7 | 40.1 | 86.4 |
Smoke pack years |
Range in months | 0.15–154 | 1–230 | 32–224 | 7–80 | − |
Mean | 41.9 | 47.3 | 64.0 | 46.5 | − |
Next, using Kaplan–Meier estimates, we show that PATQUANT allows good prognostic stratification of patients when implemented as the IASLC‐proposed grading system (accounting for less than 20% high‐risk patterns in lower grade ranks), shown in Figure 6D , but not reaching statistical significance for differentiating G1 ... |
As our tool allows for fully quantitative assessment of pattern abundance, we conducted a systematic evaluation of different cut‐offs (Figures S21–S23) for OS, CSS, and PFS endpoints, showing that higher percentages of high‐risk patterns can still be tolerated in G1/G2, with purely dominant pattern‐based grading (WHO) ... |
### Novel Approaches to Grading Allow for Fine‐Granular Patient Stratification |
Finally, we propose and explore two new approaches to LUAD grading and demonstrate that these enable more fine‐grained stratification of patients with resectable LUAD based on the risk of adverse outcomes (Figure 7). |
First, given our previous observation of the more favorable risk profile of the complex‐glandular pattern, compared to classical high‐risk patterns (micropapillary, solid), we test a four‐tier grading system (Figure 7A). We show that in our four‐cohort merged patient collective (excluding TCGA given biases mentioned ea... |
Second, inspired by Gleason grading in prostate cancer, we investigate whether the combination of primary and secondary patterns might offer even better patient stratification. We develop a scoring system that results in five prognostic groups (Figure 7D). Notably, this new scoring system enables more fine‐grained pati... |
### Analysis of Prognostic Value in Stage I and pT1 Tumors |
In addition, we evaluated the prognostic value of the WHO/IASLC grading systems, the new four‐tier grading system, and the new five‐tier scoring system (as implemented in PATQUANT) in patients with Stage I and pT1 tumors. Overall survival (OS) was selected as the primary endpoint for this subcohort, as fewer events wer... |
Notably, the new four‐tier prognostic system maintains strong prognostic capabilities when applied to the pT1 (Figure S24A) and Stage I (Figure S24B; uni‐ and multivariate analysis in Table S10) subgroups. Additionally, the new five‐tier prognostic scoring shows prognostic performance comparable to that of the all‐stag... |
## Discussion |
In our study, we developed PATQUANT, a powerful AI tool for fully automated and quantitative histological assessment of LUAD patterns in H&E‐stained WSIs. We demonstrated that the developed tool outperformed 8 out of 13 experienced pathologists in large interobserver agreement study and performed exceptionally well as ... |
LUAD patterns are a main component of the modern grading systems, with a clearly recognized prognostic value [2, 3]. Although the WHO classification still recommends using five common patterns (lepidic, acinar, papillary, micropapillary, and solid) and a three‐tier grading based on the dominant pattern [2], there is cl... |
In our study, for the annotations of the training dataset, we used Moreira et al. [3] description of complex glandular pattern as a reference. Importantly, in our highly representative, large, multi‐institutional, and multi‐national cohort of patients with resectable LUAD, we made an important observation concerning co... |
In our large interobserver agreement study, we further show that pathologists and the AI tool detection exhibit high levels of agreement in detecting this pattern, allowing for robust recognition (Figure 3B,C). Thus, pathologists reached consensus on the complex glandular pattern being dominant in 55/62 ROIs (Figure 3B... |
Moreover, we conducted a detailed analysis using different thresholds for high‐grade patterns in the IASLC grading approach (Figures S21–S23) and concluded that predominant‐pattern‐based grading (the WHO‐recommended approach) provides better prognostic stratification of patients, at least with the developed tool that w... |
We propose and implement a new scoring system using PATQUANT, inspired by Gleason grading, which accounts for dominant and secondary patterns (Figure 7). This system should be regarded as a further refinement of the scoring concept introduced by Sica et al. [28]. Different to Sica et al. and similar to Gleason grading,... |
All tested grading implementations (IASLC, WHO, new four‐tier grading system, new five‐tier scoring system) show the prognostic performance in Stage I and pT1 tumor similar to the results from the all‐stage cohort. |
A newly developed PATQUANT represents a powerful tool for pattern quantification and downstream grading applications. Several other AI‐based tools for pattern analysis have been reported. Lami et al. [17] developed a classification tool for patch‐level detection of patterns. This tool has several limitations, such as i... |
PATQUANT has several advantages over ANORAK. We used a manual annotation area at least 100 times larger, derived from 177 annotated slides across multiple departments, compared to 49 WSIs in Pan et al. [21]. This difference is likely reflected in the significantly lower segmentation accuracy for ANORAK, recognized by a... |
Our study is not without limitations. We used retrospective patient cohorts for validation, and therefore, prospective validation is warranted. The superior prognostic performance of our new scoring systems should be confirmed using independent patient cohorts. Our prognostic cohorts are underpowered for lepidic‐predom... |
## Conclusion |
In this study, we developed a powerful, fully automated AI tool for LUAD pattern detection and quantification (PATQUANT). PATQUANT outperformed 8 out of 13 experienced pathologists in a comprehensive interobserver agreement analysis. Using a large multi‐national patient cohort, we validated the prognostic capabilities ... |
## Materials and Methods |
### Patient Cohorts and Digitization |
Five patient cohorts with resectable LUAD were included in this study (Figure 1D, detailed clinicopathological characteristics are presented in Table 1): The Cancer Genome Atlas (TCGA; international) LUAD cohort consisting of patients from multiple institutions (n = 36), the National Lung Screening Trial (NLST; United ... |
### Training Dataset and Manual Annotations |
The WSIs of 177 patients from TCGA cohort were used for training purposes, selected manually to represent a broad range of morphological patterns and significant inter‐institutional/lab heterogeneity. On all slides, several (1–5) representative regions were selected for manual annotations, which were performed by two b... |
### Algorithm Development |
Annotated regions were extracted in the form of patches of size 512 × 512 px at MPP 1.0 (roughly 10x objective magnification) with an overlap of 200 px. Annotations were exported as masks with class coding, including class “0” for non‐annotated regions. Patches with less than 5% of annotated area were ignored. Backgrou... |
### Model Inference and WSI Pipeline |
The model is applied to regions detected as tumor and is integrated with a non‐small cell lung cancer (LUAD, lung squamous cell carcinoma) detection tool that was developed and extensively validated previously [14]. The entire pipeline consists of three algorithms: tissue detection (1x magnification analysis), tumor de... |
### Test Dataset: Formal Validation |
One independent dataset (UKK) with 20 LUAD resection cases (WSI n = 20; H&E‐stained), representing a broad spectrum of patterns, was manually annotated with high precision (using principles similar to those outlined above for the training dataset). This dataset was used for pixel‐level formal statistical validation of ... |
### Test Dataset: Interobserver Agreement Analysis |
A dataset consisting of large regions of interest (ROIs) from multiple cases across three international cohorts of resection cases (Germany: UKK, Japan: KAMEDA, USA: cohort of the National Cancer Institute's Clinical Proteomic Tumor Analysis Consortium/CPTAC) was created (n = 283). This dataset is representative of thr... |
### Pathologists |
Thirteen pathologists specializing in lung cancer from five countries were included in the study. All of them provided their pattern analysis for 283 ROIs of interobserver agreement analysis dataset. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.