Artificial intelligence grading of breast cancer: a promising method to refine prognostic classification for management precision.

AIM: Artificial intelligence (AI)-based breast cancer grading may help to overcome perceived limitations of human assessment. Here, the potential value of AI grade was evaluated at the molecular level and in predicting patient outcome. METHODS AND RESULTS: A supervised convolutional neural network (CNN) model was trained on images of 612 breast cancers from The Cancer Genome Atlas (TCGA). The test set, obtained from the Cooperative Human Tissue Network (CHTN), comprised 1058 cancers with corresponding survival data. Upon reversal, a CNN was trained from images of 1537 CHTN cancers and tested on 397 TCGA cancers. In TCGA, mRNA models were trained using AI grade and Nottingham grade (NG) as labels. Performance of mRNA models in predicting patient outcome was evaluated using data from 1807 cancers from the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) cohort. In selecting images for training, nucleolar prominence determined high- versus low-grade cancer cells. In CHTN, NG corresponded to significant survival stratification in stages 1, 2 and 3 cancers, while AI grade showed significance in stages 1 and 2 and borderline in stage 3 tumours. In METABRIC, the mRNA model trained from AI grade was not significantly different to the NG-based model. The gene which best described AI grade was TRIP13, a gene involved with mitotic spindle assembly. CONCLUSION: An AI grade trained from the morphologically distinctive feature of nucleolar prominence could transmit significant patient outcome information across three independent patient cohorts. AI grade shows promise in gene discovery and for second opinions.

The nucleolar-related protein Dyskerin pseudouridine synthase 1 (DKC1) predicts poor prognosis in breast cancer.

BACKGROUND: Hypertrophy of the nucleolus is a distinctive cytological feature of malignant cells and corresponds to aggressive behaviour. This study aimed to identify the key gene associated with nucleolar prominence (NP) in breast cancer (BC) and determine its prognostic significance. METHODS: From The Cancer Genome Atlas (TCGA) cohort, digital whole slide images identified cancers having NP served as label and an information theory algorithm was applied to find which mRNA gene best explained NP. Dyskerin Pseudouridine Synthase 1 (DKC1) was identified. DKC1 expression was assessed using mRNA data of Molecular Taxonomy of Breast Cancer International Consortium (METABRIC, n = 1980) and TCGA (n = 855). DKC1 protein expression was assessed using immunohistochemistry in Nottingham BC cohort (n = 943). RESULTS: Nuclear and nucleolar expressions of DKC1 protein were significantly associated with higher tumour grade (p < 0.0001), high nucleolar score (p < 0.001) and poor Nottingham Prognostic Index (p < 0.0001). High DKC1 expression was associated with shorter BC-specific survival (BCSS). In multivariate analysis, DKC1 mRNA and protein expressions were independent risk factors for BCSS (p < 0.01). CONCLUSION: DKC1 expression is strongly correlated with NP and its overexpression in BC is associated with unfavourable clinicopathological characteristics and poor outcome. This has been a detailed example in the correlation of phenotype with genotype.

Prognostic significance of nucleolar assessment in invasive breast cancer.

AIMS: Nucleolar morphometric features have a potential role in the assessment of the aggressiveness of many cancers. However, the role of nucleoli in invasive breast cancer (BC) is still unclear. The aims of this study were to investigate the optimal method for scoring nucleoli in IBC and their prognostic significance, and to refine the grading of breast cancer (BC) by incorporating nucleolar score. METHODS AND RESULTS: Digital images acquired from haematoxylin and eosin-stained sections from a large BC cohort were divided into training (n = 400) and validation (n = 1200) sets for use in this study. Four different assessment methods were evaluated in the training set to identify the optimal method associated with the best performance and significant prognostic value. These were: (i) a modified Helpap method; (ii) counting prominent nucleoli (size ≥2.5 µm) in 10 field views (FVs); (iii) counting prominent nucleoli in five FVs; and (iv) counting prominent nucleoli in one FV. The optimal method was applied to the validation set and to an external validation set, i.e. data from The Cancer Genome Atlas (n = 743). Scoring prominent nucleoli in five FVs showed the highest interobserver concordance rate (intraclass correlation coefficient of 0.8) and a significant association with BC-specific survival (P < 0.0001). A high nucleolar score was associated with younger age, larger tumour size, and higher grade. Incorporation of nucleolar score in the Nottingham grading system resulted in a higher significant association with survival than the conventional grade. CONCLUSIONS: Quantification of nucleolar prominence in five FVs is a cost-efficient and reproducible morphological feature that can predict BC behaviour and can provide an alternative to pleomorphism to improve BC grading performance.

Visual histological assessment of morphological features reflects the underlying molecular profile in invasive breast cancer: a morphomolecular study.

AIMS: Tumour genotype and phenotype are related and can predict outcome. In this study, we hypothesised that the visual assessment of breast cancer (BC) morphological features can provide valuable insight into underlying molecular profiles. METHODS AND RESULTS: The Cancer Genome Atlas (TCGA) BC cohort was used (n = 743) and morphological features, including Nottingham grade and its components and nucleolar prominence, were assessed utilising whole-slide images (WSIs). Two independent scores were assigned, and discordant cases were utilised to represent cases with intermediate morphological features. Differentially expressed genes (DEGs) were identified for each feature, compared among concordant/discordant cases and tested for specific pathways. Concordant grading was observed in 467 of 743 (63%) of cases. Among concordant case groups, eight common DEGs (UGT8, DDC, RGR, RLBP1, SPRR1B, CXorf49B, PSAPL1 and SPRR2G) were associated with overall tumour grade and its components. These genes are related mainly to cellular proliferation, differentiation and metabolism. The number of DEGs in cases with discordant grading was larger than those identified in concordant cases. The largest number of DEGs was observed in discordant grade 1:3 cases (n = 1185). DEGs were identified for each discordant component. Some DEGs were uniquely associated with well-defined specific morphological features, whereas expression/co-expression of other genes was identified across multiple features and underlined intermediate morphological features. CONCLUSION: Morphological features are probably related to distinct underlying molecular profiles that drive both morphology and behaviour. This study provides further evidence to support the use of image-based analysis of WSIs, including artificial intelligence algorithms, to predict tumour molecular profiles and outcome.

Breast cancer histologic grading using digital microscopy: concordance and outcome association.

AIMS: Virtual microscopy utilising digital whole slide imaging (WSI) is increasingly used in breast pathology. Histologic grade is one of the strongest prognostic factors in breast cancer (BC). This study aims at investigating the agreement between BC grading using traditional light microscopy (LM) and digital WSI with consideration of reproducibility and impact on outcome prediction. METHODS: A large (n=1675) well-characterised cohort of BC originally graded by LM was re-graded using WSI. Two separate virtual-based grading sessions (V1 and V2) were performed with a 3-month washout period. Outcome was assessed using BC-specific and distant metastasis-free survival. RESULTS: The concordance between LM grading and WSI was strong (LM/WSI Cramer's V: V1=0.576, and V2=0.579). The agreement regarding grade components was as follows: tubule formation=0.538, pleomorphism=0.422 and mitosis=0.514. Greatest discordance was observed between adjacent grades, whereas high/low grade discordance was uncommon (1.5%). The intraobserver agreement for the two WSI sessions was substantial for grade (V1/V2 Cramer's V=0.676; kappa=0.648) and grade components (Cramer's V T=0.628, p=0.573 and M=0.580). Grading using both platforms showed strong association with outcome (all p values <0.001). Although mitotic scores assessed using both platforms were strongly associated with outcome, WSI tends to underestimate mitotic counts. CONCLUSIONS: Virtual microscopy is a reliable and reproducible method for assessing BC histologic grade. Regardless of the observer or assessment platform, histologic grade is a significant predictor of outcome. Continuing advances in imaging technology could potentially provide improved performance of WSI BC grading and in particular mitotic count assessment.

The Advantage of Discordance: An Example Using the Highly Subjective Nuclear Grading of Breast Cancer.

Discordance among multiple assessments has been a reason to criticize a biomarker. But, if different assessments are all relevant, the meaning of discordance requires explanation. As an example, for 1085 breast cancers, a low (score 1), intermediate (score 2) or high nuclear grade (NG) (score 3) was assigned in years 2013, 2015, 2016. Year apart readings allowed for memory lapse of prior readings. For each cancer, scores for NG2013, NG2015, NG2016 were added together to yield sum score nuclear grade (SSNG) with range 3 to 9. SSNG was used to find if discrepancy between NG readings carried information for patient outcome. Discrepancies were inherent with SSNG=4, 5, 7 or 8. Time-dependent receiver operator curves were central for evaluating discordance as related to patient outcome. Area under curves for SSNG, and the component NGs, in stage 1, stage 2, and stage 3 cancers were, respectively: SSNG: 70, 68, 75; NG2013: 70, 63, 71; NG2015: 67, 65, 74; and NG2016: 65, 66, 68. The area under curves of SSNG was not significantly lower than any of the components from which it was derived. This is despite discordant readings having been incorporated into SSNG. Among the 3 readings, 50.1% were discordant, yet only 2.1% were low/high discrepancy. Concordance in high-grade assignment (SSNG=9) corresponded to poor prognosis. If morphologic features are midway between 2 predefined levels it is sensible that separate readings will be distributed between adjacent levels. Shown has been how an "in-between" level helps predict survival then discordance discovery offers classification. Discordance discovery can conceivably be embraced for real-world applications.