Classification of Non-Small Cell Lung Cancer Stage using a Convolutional Neural Network
Authors
Gade, Josefine Dam ; Hald, Line Sofie
Term
4. term
Publication year
2019
Pages
80
Abstract
Lungekræft er den største årsag til kræftdødsfald. Den mest almindelige form, ikke-småcellet lungekræft (NSCLC), har fire undertyper, og primærtumorer stadieinddeles fra I til IV. Stadieinddeling foretages typisk af patologer på histopatologiske snit. Fordi prøverne rummer mange celler og farveintensiteten kan variere, kan forskellige eksperter nå forskellige vurderinger. Konvolutionelle neurale netværk (CNN’er)—dybdelæringsmodeller til billedanalyse—er blevet brugt til at klassificere lungekræft i undertyper, men deres evne til at bestemme tumorstadie ud fra billeder er ikke undersøgt. Dette studie afprøver, om en CNN kan forudsige NSCLC-stadie fra histopatologi. Vi brugte data fra bronchus- og lunge-delen af The Cancer Genome Atlas (TCGA). Efter dataudvælgelse og forbehandling for at øge ensartethed finjusterede vi en fortrænet AlexNet-model (en udbredt CNN). På grund af bekymring for støj i labels gennemførte vi fire eksperimenter. I hvert eksperiment anvendte vi tre WSI-klassifikationsmetoder (whole-slide images) til at evaluere de forudsagte labels. Den trænede model blev evalueret både på små billedudsnit (patches) og på hele snit-niveau. På tværs af de fire eksperimenter var nøjagtigheden på patch-niveau 0,52 ± 0,04. De bedste eksperimenter opnåede 0,56 i alle tre tilgange. Ved WSI-baseret klassifikation nåede to af tilgangene 0,56 ± 0,01. Resultaterne dokumenterer den indledende ydeevne for CNN-baseret stadieinddeling af NSCLC fra histopatologi i dette datasæt og setup.
Lung cancer is the leading cause of cancer deaths. The most common form, non-small cell lung cancer (NSCLC), has four subtypes, and primary tumors are staged from I to IV. Staging is usually done by pathologists who examine histopathology slides. Because samples contain many cells and staining intensity varies, different experts may reach different judgments. Convolutional neural networks (CNNs)—deep learning models for image analysis—have been used to classify lung cancer subtypes, but their ability to determine tumor stage from images has not been studied. This study investigates whether a CNN can predict NSCLC stage from histopathology. We used bronchus and lung data from The Cancer Genome Atlas (TCGA). After selecting and preprocessing data to improve consistency, we fine-tuned a pretrained AlexNet model (a widely used CNN). Due to concerns about noisy labels, we ran four experiments. In each, we applied three whole-slide image (WSI) classification approaches to evaluate predicted labels. The trained model was evaluated both on small image patches and at the whole-slide level. Across the four experiments, patch-level accuracy was 0.52 ± 0.04. The best experiments achieved 0.56 accuracy in all three approaches. For WSI-based classification, two of the approaches reached 0.56 ± 0.01. These results document initial performance for CNN-based staging of NSCLC from histopathology within this dataset and setup.
[This abstract was generated with the help of AI]
Documents