Quantitative wood anatomy (QWA) has proven to be a powerful method for extracting relevant environmental information from tree-rings. Although classical image-analysis tools such as ROXAS have greatly improved and facilitated measurements of anatomical features, producing QWA datasets remains challenging and time-consuming. In recent years, deep learning techniques have drastically improved the performance of most computer vision tasks. We, therefore, investigate three different deep learning models (U-Net, Mask-RCNN, Panoptic Deeplab) to improve the main bottleneck, cell detection. Therefore, we create a Conifer Lumen Segmentation (CoLuS) dataset for training and evaluation. It consists of manual outlines of each cell lumen from anatomical images of several conifer species that cover a wide range of sample qualities. We furthermore apply our deep learning model to a previously published high-quality QWA chronology from Northern Finland to compare the warm-season (AMJJAS) temperature reconstruction skill of our deep learning method with that of the current ROXAS implementation, which is based on classical image analysis. Based on our evaluation dataset we show improvements of 7.6% and 8.1% for our best performing deep learning model (U-Net) for the computer vision metrics mean Intersection over Union (mIoU) and Panoptic Quality (PQ) compared to automatic ROXAS segmentation, in addition to being much faster. Furthermore, U-Net reduces the percentage error compared to automatic ROXAS analysis - which tends to systematically underestimate lumen area - by 57.8% for lumen area, 63.2% for average cell wall thickness, and 54.1% for cell count. In addition, we show higher performance for the U-Net compared to the Mask-RCNN previously used for tree cell segmentation. These improvements are independent of sample quality. For the Northern Finland QWA chronology, our U-Net model matches or outperforms ROXAS with and without manual post-processing, showing a common signal (Rbar) of 0.72 and a AMJJAS temperature correlation of 0.81 for maximum radial cell wall thickness. A clear improvement is especially visible for the anatomical latewood density, likely due to the better detection of small cell lumina. Our results demonstrate the potential of deep learning for higher-quality segmentation with lower manual post-processing time, saving weeks to months of tedious work without compromising data quality. We thus plan to implement deep learning in a future version of ROXAS.