Positron emission tomography is a nuclear imaging technique well known for its use in oncology for cancer diagnosis and staging. A PET scanner is a complex machine which comprises photodetectors placed in a ring configuration that detect gamma photons generated through annihilation between an electron and a positron. The accuracy with which the gamma photons are detected determines the quality of the extracted information, which is then further analysed through image reconstruction algorithms. Therefore, the design and optimization of photodetectors and their readout electronics is very important for the advancements of PET scanners. For more than a decade, silicon photomultipliers have been researched extensively and became the photodetectors of choice for this application. High performance readout electronics is essential in order to measure data with high precision and various readout schemes have been proposed over the years for PET photodetector modules. As PET scanners are complex systems, research is dedicated individually to different parts. This thesis focuses on the design of integrated readout electronics for time-of-flight PET application. Consequently, the readout of three SPAD-based sensors was designed. The first sensor, Blumino represents the first fully integrated analog silicon photomultiplier with on-chip discrimination and time conversion. The design was implemented in 350 nm CMOS technology node. The chip serves as a prototype for future fully integrated A-SiPMs as PET photodetectors due to its simplicity and compactness. The second sensor, Blueberry, advances the previous design by exploring the benefits of multi-digital silicon photomultipliers and 3D integration. The sensor was designed in a 3D-stacked FSI CMOS technology node enabling features such as: improved spatial and timing resolution (a TDC design with on-chip error correction algorithm of 15 ps LSB). The third sensor, Smarty, is an on-chip fully reconfigurable neural network with 10 TDCs designed in 16 nm FinFET technology. The chip is capable of executing 363 MOPS and was designed for pre-processing and data compression at the sensor level. Various neural network configurations were explored and trained using genetic algorithms. The architecture was proven to be viable for reconstructing radioactive source positions in a coincidence setup.